Installation and deployment

This page takes you from an empty host to a running instance you can leave alone, and explains enough of the compose file that you can change it without breaking it. After reading it you should be able to size a machine, decide which of the optional containers you actually want, put TLS in front of the API, upgrade without losing data, and run the whole thing without Docker if you prefer.

If you only want the shortest path to a working instance, read Quick start first and come back here when something needs changing.

Which path

There are two ways to install this. Pick one before reading further.

	Docker Compose	By hand
Who it is for	Almost everyone, production included	Environments where Docker is not an option, or you are changing the code
You install on the host	Docker, and nothing else	PostgreSQL 17 + TimescaleDB, Redis 8, Python 3.12, uv — plus Node 22 if you build the console
Time to a running stack	However long the pull takes	Half an hour and up, half of it on TimescaleDB
Upgrading	./dtkctl upgrade, or by hand git pull, pull, migrate, up -d	git pull, uv sync, migrate, restart the systemd units
Resource ceilings, read-only rootfs, dropped capabilities, non-root user	Already in the compose file	You rebuild them yourself with systemd
When you ask me about a problem	I can reproduce it exactly	Your box is one of a kind and I am guessing
Start reading at	First install	Running without Docker for development, A production install without Docker for a server

Docker is the recommendation, not because containers are fashionable but because everything this stack is fussy about is already pinned in the compose file: the database has to be Postgres 17 with the TimescaleDB extension (a plain postgres:17 image will not do), the browser image pins CloakBrowser to one commit, the four services are sequenced by health checks, and every container’s memory, CPU and PID ceilings and read-only root filesystem are written down. By hand you rebuild all of that yourself — A production install without Docker lists what you are putting back, and it is worth reading before you choose.

Contents

Before you start

• Installing and managing it with the script — the guided install, and the operations menu afterwards
• What you need — sizing, disk, and the two things to know before pointing this at your own database
• What the shipped defaults are written for · Three sizes · Changing the limits · When transcription and the browser share a host
• Network preparation in mainland China — switch your mirrors first; without them the install usually dies on the image pull
• The stack at a glance — how many containers, and which talks to which

With Docker

• First install — empty directory to a console you can log in to
• The compose file, service by service — read this before editing it
• The optional profiles — whether you want browser, downloader and transcriber
• Building the browser image: the CloakBrowser pin — this image is not published; you build it
• Environment variables — including the two ways Compose reads .env, which catches everyone once
• Volumes · Ports and binding · Behind a reverse proxy
• Verifying an install — run these; do not judge it by whether the page loads
• Upgrading · Scaling: more workers

By hand

• Running without Docker — how the project is developed: uv for dependencies, containers still for Postgres and Redis
• A production install without Docker — all of it by hand, systemd units included, verified on a clean Ubuntu 24.04
• Running the worker on its own

When something is wrong

• Starting over — stopping with the data intact, and stopping with it gone
• Where to go next

What you need

Requirement	Why
Docker Engine with Compose v2 (v2.24 or newer)	v2.24 is where Compose added the long form env_file: - path: … required: false, which docker/compose.yml uses, and COMPOSE_ENV_FILES, which this page suggests. Nothing in the repository enforces the floor, so it is a property of the features used, not a declared minimum
Linux, macOS or WSL2 host	Nothing in the stack is platform-specific, but the browser container wants a lot of memory and is happiest on Linux
2 GiB RAM for the core stack, 4 GiB to be safe	2 GiB is enough in practice — measured usage at rest is about 295 MiB across the four core containers (see the table below) — but the container ceilings sum to 3.5 GiB, so 4 GiB is the safe figure
~8 GiB RAM if you run the browser profile	The browser container is capped at 4 GiB and genuinely uses it
~1 GiB more RAM and two cores if you run the transcriber profile	The models stay loaded, about 750 MiB at rest; a job keeps two cores busy while it runs, and only then
~15 GB of free disk	Roughly 4.5 GB of images for the core stack and ~6.3 GB with the browser container, before volumes and build cache; then Postgres and, if you enable media downloads, the media volume (capped at 2 GiB by default). Transcription adds a ~1.1 GB image and ~500 MB of models. 15 GB free is comfortable
Outbound HTTPS	The worker and the browser container talk to the platforms, usually through proxies you supply

You do not need to install Postgres, Redis, Python, Node or a browser on the host. The compose file brings all of them, pinned.

Two things about the database are worth knowing before you swap it for one you already run:

• The schema requires the TimescaleDB extension. Request logs, identity events and content snapshots are hypertables, and the first migration checks pg_available_extensions and refuses with an explanation rather than creating half a schema. The compose file uses timescale/timescaledb-ha:pg17; a stock postgres:17 image will not do.
• Nothing in the repository ships a default password or key. The entrypoint exits 78 (EX_CONFIG) if DTK_SECRET_KEY is missing or shorter than 32 characters, and the redis container exits if REDIS_PASSWORD is empty. That is deliberate: a key that is identical on every install is not encryption.

Resource ceilings and measured usage, both read off docker/compose.yml (the measurements in its comments were taken on 2026-09-08):

Service	mem_limit	cpus	pids_limit	Measured at rest
postgres	2g	—	256	112 MiB
redis	512m	—	128	8 MiB
api	512m	2.0	256	104 MiB
worker	512m	2.0	256	71 MiB
browser-rpc	4g	4.0	1024	2.57 GiB warm, 629% CPU
downloader	512m	2.0	128	streaming buffers only
transcriber	4g	2.0	256	~750 MiB with the models loaded; ~0.9 GiB transcribing a 3-minute clip

The transcriber row was measured later, on 2026-10-16, with the quantized models from Preparing the FunASR ONNX models.

The limits exist so one overloaded component cannot destabilise the rest. postgres is the exception with no CPU cap and a loose memory cap, because a tight ceiling on Postgres degrades correctness rather than protecting anything — its 2g is there to stop a runaway query taking the host, not to be a budget it has to live inside.

What the shipped defaults are written for

Add up the ceilings in the table above: about 8 GiB of memory, and browser-rpc asks for 4 CPUs. So docker/compose.yml describes a 4 vCPU / 8 GiB machine — on one of those, up -d needs no changes at all.

Ceilings are not reservations, so day-to-day usage sits far below them. But on a smaller box this is not “slower”, it is “will not start”:

Error response from daemon: range of CPUs is from 0.01 to 2.00,
as there are only 2 CPUs available

Docker refuses to start a container asking for more CPUs than the host has. It is a hard error, not a cap that degrades quietly.

Three sizes

Size	Spec	What you get	Compose changes
No browser	1 vCPU / 2 GiB	Manual cookie import for Douyin and TikTok; only Bilibili, which needs no browser, is minted automatically. The four core containers idle at about 295 MiB	None — just leave --profile browser off
Full, minimum	2 vCPU / 4 GiB plus swap	Everything	Yes, see the override below
Full, recommended	4 vCPU / 8 GiB	Everything	None

The minimum row is measured rather than estimated. A 2 vCPU / 3.8 GiB VPS running the whole stack with the browser profile on:

	Idle	Peak during a mint
browser-rpc	234 MiB	845 MiB
api	115 MiB	128 MiB
postgres	80 MiB	88 MiB
worker / redis / downloader	82 / 4 / 4 MiB	barely move
Host total	1.0 GiB	1.1 GiB

What sets the floor is the mint peak, not the idle figure: Chromium spikes while it opens a context, and that is the same moment Postgres is holding its buffers. Give a box this size swap — 2 GiB with vm.swappiness=10 is enough. It is not there to be used; it is there so that peak becomes a slow second instead of the kernel picking a process to kill, which is usually Postgres rather than the one that actually grew.

Changing the limits

Do not edit docker/compose.yml. It describes a normal machine, and the way yours differs should be visible in a file of its own. Write compose.host.yml at the repository root:

services:
  postgres:    { mem_limit: 1g,    memswap_limit: 2g }
  redis:       { mem_limit: 320m,  memswap_limit: 512m }
  api:         { mem_limit: 448m,  memswap_limit: 768m,  cpus: 1.5 }
  worker:      { mem_limit: 448m,  memswap_limit: 768m }
  browser-rpc: { mem_limit: 1200m, memswap_limit: 2400m, cpus: 1.5 }
  downloader:  { mem_limit: 192m,  memswap_limit: 384m,  cpus: 1.0 }

Then pass it on every command:

COMPOSE_ENV_FILES=.env docker compose -p dtk \
  -f docker/compose.yml -f compose.host.yml --profile browser up -d

Three arguments, none of them optional, all of them easy to forget — which is what a wrapper is for. (The guided script writes this file, and the compose.host.yml above it, with the numbers worked out for your machine. What follows is what it does, and how to write it yourself.)

cat > dtkctl <<'EOF'
#!/usr/bin/env bash
set -euo pipefail
cd "$(dirname "$0")"
export COMPOSE_ENV_FILES=.env
exec docker compose -p dtk -f docker/compose.yml -f compose.host.yml \
  --profile browser --profile downloader "$@"
EOF
chmod +x dtkctl
./dtkctl up -d
./dtkctl logs -f api

Knobs worth turning, most effective first:

• Give browser-rpc fewer cpus. Signing and minting are both off the request path, so it is exactly the service that should yield when the machine is busy. 1.5 rather than 2.0 leaves api a core to answer on.
• DTK_BROWSER_WARM_CONTEXTS (default 1). Each resident signing page costs roughly 300 MiB. Leave it at 1 on a small host.
• DTK_DOWNLOADER_WORKERS / DTK_DOWNLOADER_ITEM_WORKERS (4 each). The concurrent-transfer ceiling is their product. Drop both to 2 on one core.
• pool.target_size (a console setting). A bigger pool means more mints and more rows in Postgres. 8 is a sensible number on a small box.
• DTK_TRANSCRIBER_THREADS (default 2). ONNX threads for one transcription. Keep it at or under the transcriber’s cpus; more threads than cores only adds switching. If the browser runs on the same host, read the next section.

When transcription and the browser share a host

browser-rpc and transcriber are both CPU-bound background work, and their ceilings — 4 cores and 2 — are caps, not reservations. On a host without that many idle cores they take turns on the same ones, and the transcription is what you notice, because somebody is waiting for its text.

Measured on 2026-10-16 with the transcriber at its shipped two cores:

Situation	Audio	Time
Transcriber alone	3 min 13 s	3.7 s
Four busy threads pinned to the same two cores	3 min 13 s	15.7 s
A busy development host while browser-rpc minted an identity	21 s	31 s and 50+ s, against under a second once the mint had finished

What follows from that:

• A slow transcription during a mint is expected, and it finishes. Mints happen at startup, while the pool fills, and whenever an identity retires. DTK_TRANSCRIBER_TIMEOUT_SECONDS (3600 by default) is far above any of this.

• ?wait= may come back running. It waits at most 30 seconds; after that, poll the transcription or its task rather than calling again with refresh.

• Let the pool fill before a batch. A dozen transcriptions queued right after up -d all run while the browser is busiest.

• On a host with six or more cores, give them different cores. cpuset in compose.host.yml keeps them apart entirely:

  services:
    browser-rpc: { cpuset: "0-3" }
    transcriber: { cpuset: "4-5" }

• On a smaller host, shrink the browser rather than the transcriber. The first knob above already says why: minting is the work that can wait.

The guided script writes transcriber: { mem_limit: 1536m, memswap_limit: 1536m } into compose.host.yml on a host under about 7.8 GiB, so the 4g ceiling does not add to a total the machine does not have. It does not set cpuset: which cores are free depends on what else the host runs.

Installing and managing it with the script

There is a guided script that walks the Docker route above for you, and then manages the instance afterwards. It exists in both languages — see install/README.md.

curl -fsSL https://raw.githubusercontent.com/Evil0ctal/Douyin_TikTok_Download_API/main/install/install.sh -o install.sh
less install.sh    # read it before you run it
bash install.sh

The first run works out the distribution, checks for Docker (and offers to install it), sizes the resource ceilings to the cores and memory Docker can use (on Docker Desktop or OrbStack, the VM’s), then asks seven questions: where, which address and port, the browser container, media downloads, offline transcription, published images or a build. It clones main and checks out the latest release’s tag, detached — DTK_VERSION names another release, DTK_REPO_URL or DTK_GH_PROXY a mirror to clone from, which it asks about first — and writes that version into .env as DTK_IMAGE_TAG, not latest. It will not start over the volumes of an earlier install with the same project name: their database opens only with that install’s .env, so it stops and says what to do instead. Say yes to transcription and it also asks where the models are: a directory that already holds them is used as it is, and an empty one is offered a download — about 500 MB from ModelScope, pinned to the commits that Preparing the FunASR ONNX models verified, with every file checked against its SHA-256. With the browser on, the image installs the CloakBrowser pin that docker/compose.yml names; the script does not copy it into .env, so it moves with the checkout on every upgrade. It writes three things, and they are exactly the three this page goes on to teach you to write by hand:

What it writes	The section here
.env, secrets from openssl rand, mode 0600	First install, step 2
compose.host.yml, ceilings scaled to this host	Changing the limits
dtkctl, carrying the project name, both compose files and COMPOSE_ENV_FILES	the same section

Run it again and it finds the install — by asking Docker, so it does not matter where you put it — and opens a menu instead of installing. It looks with --yes and with DTK_INSTALL_DIR set too; under --yes, or without a terminal, it changes nothing and points at ./dtkctl upgrade. The menu:

• Status — versions, containers, disk, and a comparison against the latest GitHub release
• Upgrade — moves the checkout to the target release’s tag (an install built from source follows main), checks the new compose configuration, pulls the release (or rebuilds the images, on an install built from source), rebuilds the browser image, takes a backup, migrates, and only then writes the version into .env and restarts. A failure before the restart puts the checkout and .env back, with the old version still running
• Manage — passwords, an extra administrator, account list, backup, restore, self check, logs, restart, any runtime setting, disk cleanup, and transcription on or off. Turning it on runs the same model step as the install, pulls or builds the image, and recreates the stack; turning it off removes the container and leaves the models where they are. Either way only the one --profile flag in dtkctl changes, so anything else you edited in that file survives. Restore takes an archive’s name from the list, and like every restore it adds what is missing without overwriting anything (Operations)
• Stop or remove — three levels: stop (the containers are kept, so the install is still found; ./dtkctl up -d starts it), stop and delete the volumes, or that plus the install directory. The last two require typing delete, not a y/n, and say how to copy the backups out first

bash install.sh --manage goes straight to that menu, and the upgrade has a command of its own that asks nothing, run in the install directory:

./dtkctl upgrade                    # the latest release
./dtkctl upgrade 5.1.6              # or that one; also v5.1.6, 5.1 (the newest 5.1.x), sha-<commit>
./dtkctl upgrade 5.1.6 --no-backup  # without the backup it takes first

That is the whole upgrade - the version in .env is the script’s to keep, not yours. After moving the checkout to the release’s tag it hands over to the copy of the script that came with it, so the steps it runs are always the new version’s. Then, in this order: Compose’s check of the new configuration, the pull or build, the browser image, a backup with dtk backup create into the backup volume (only if the API is running; a failed backup stops the upgrade), the migrations, and only after those the new DTK_IMAGE_TAG in .env and the restart. Any failure before the restart, or Ctrl-C, puts the checkout and .env back as they were (Put back: ...) and leaves the running version alone, and running it again finishes an upgrade that stopped half-way. A release that does not have the database’s latest migration is refused before anything is pulled: migrations do not run backwards, and the way back to it is a new install of that version with a backup made on it restored. An install made before 5.1.6 gets the command the first time it is upgraded: from the menu, or once with git pull && bash install/install.sh --upgrade in its directory.

An upgrade also takes out a CLOAKBROWSER_COMMIT line that this project wrote into .env itself (CloakBrowser 0.5.10 from an old .env.example, or 0.6.3 from the installers of 5.1.4 to 5.1.6), leaving a comment in its place, so the browser image follows the checkout from then on. A revision you chose yourself is kept, with a warning that the checkout’s browser dependencies were verified with another one. When the revision changes, the browser rebuild downloads a new Chromium and takes several minutes.

The rest of this page is still worth reading. What the script does is what follows; it just does it for you. When you want to know why it chose something, or to deviate from it, the answer is below.

Network preparation in mainland China

Skip this section unless you are installing from a machine in mainland China.

The short version: change your mirrors before you start. Almost everything this install pulls lives outside the Great Firewall — Docker Hub, PyPI, the npm registry, GitHub, Debian’s apt repositories. Without mirrors it is not slower, it usually times out on the image pull, and you conclude the project is broken.

Treat the mirrors below as leads that worked when this page was written, not as guarantees. Public mirrors have a history of disappearing — a wave of university sites dropped their Docker Hub proxies in 2024 — so each part ends with a way to check for yourself.

Docker images

On the recommended path, where you only pull prebuilt images, there are four to fetch from Docker Hub, and a fifth if you turn transcription on:

Image	Used by
evil0ctal/douyin_tiktok_download_api	api, worker and migrate, which share one image
evil0ctal/douyin_tiktok_download_api-downloader	downloader (optional profile)
evil0ctal/douyin_tiktok_download_api-transcriber	transcriber (optional profile; published from the first release that includes transcription — an older tag has no such image)
timescale/timescaledb-ha:pg17	postgres
redis:8-alpine	redis

The transcription models are not an image and do not come from Docker Hub: they are ~500 MB from ModelScope, which is usually fast from mainland China. See Preparing the FunASR ONNX models.

Mirrors go in /etc/docker/daemon.json:

{
  "registry-mirrors": ["https://<your-account-id>.mirror.aliyuncs.com"]
}

sudo systemctl daemon-reload && sudo systemctl restart docker
docker info | grep -A3 "Registry Mirrors"   # nothing printed means it did not take

Sources, most to least reliable:

• Your own cloud provider’s accelerator. Aliyun issues a per-account address from the Container Registry console; Tencent Cloud is https://mirror.ccs.tencentyun.com but only resolves inside Tencent Cloud’s network; Huawei Cloud and Volcengine are the same idea. These are the steady ones, because they are a vendor serving its own customers rather than a volunteer site.
• Public accelerators, such as https://docker.m.daocloud.io. Use one while it works. registry-mirrors takes a list and Docker tries them in order.
• When none of them work: docker pull on a machine outside, then docker save / docker load across. Or run a registry proxy on a host outside. It sounds primitive; it costs less time than cycling through accelerators that time out.

Note that the browser-rpc image is not published and has to be built locally — see Building the browser image: the CloakBrowser pin. Its build reaches Debian’s apt repositories, PyPI, GitHub (CloakBrowser installs as pip install "cloakbrowser @ git+https://github.com/…") and then downloads a browser binary. This is the most network-sensitive step in the whole install, and most of what follows exists for it. If you do not need minted Douyin or TikTok identities yet, leave the browser profile off — importing your own cookies works without it, and Bilibili identities are minted without it anyway.

System package mirrors

The bare-metal install installs PostgreSQL 17, TimescaleDB and Redis from their official apt repositories.

One Ubuntu 24.04 trap worth calling out: sources moved to the deb822 format at /etc/apt/sources.list.d/ubuntu.sources. The /etc/apt/sources.list that older guides edit no longer does anything, and editing it fails silently — you will believe you switched mirrors.

# Ubuntu 24.04, switching to the Tsinghua mirror
sudo sed -i \
  -e 's|http://archive.ubuntu.com/ubuntu|https://mirrors.tuna.tsinghua.edu.cn/ubuntu|g' \
  -e 's|http://security.ubuntu.com/ubuntu|https://mirrors.tuna.tsinghua.edu.cn/ubuntu|g' \
  -e 's|http://ports.ubuntu.com/ubuntu-ports|https://mirrors.tuna.tsinghua.edu.cn/ubuntu-ports|g' \
  /etc/apt/sources.list.d/ubuntu.sources
sudo apt-get update

The third expression is for arm64 — ARM Ubuntu uses ports.ubuntu.com rather than archive.ubuntu.com. On x86 it matches nothing and is harmless.

Common hosts, interchangeable by domain: mirrors.tuna.tsinghua.edu.cn (Tsinghua), mirrors.ustc.edu.cn (USTC), mirrors.aliyun.com (Aliyun), repo.huaweicloud.com (Huawei), mirrors.cloud.tencent.com (Tencent). On a given cloud, prefer that cloud’s own — it stays on the internal network, which is both faster and not metered.

The PostgreSQL apt repository (apt.postgresql.org) is mirrored at Tsinghua and elsewhere, on a path like https://mirrors.tuna.tsinghua.edu.cn/postgresql/repos/apt/ — check the mirror’s own help page, these paths do move.

TimescaleDB’s packages are on packagecloud.io, which as far as I know has no mirror in China. If that step stalls, either put apt behind a proxy (sudo -E apt-get … with https_proxy set) or go the Docker route — pulling timescale/timescaledb-ha:pg17 through an accelerator beats installing packages from packagecloud by a wide margin. That is one of the reasons Docker is the recommendation.

Python, pip and uv

Dependencies are managed with uv.

# uv itself: the official install script goes through GitHub, so install from PyPI when it drags
pip install uv -i https://pypi.tuna.tsinghua.edu.cn/simple

# The dependency index: uv reads this variable
export UV_DEFAULT_INDEX=https://pypi.tuna.tsinghua.edu.cn/simple
uv sync --frozen --no-dev

Put it in ~/.bashrc, or in a systemd unit’s Environment=. Older uv reads UV_INDEX_URL; setting both does no harm.

For pip itself:

pip config set global.index-url https://pypi.tuna.tsinghua.edu.cn/simple

Common PyPI mirrors: Tsinghua https://pypi.tuna.tsinghua.edu.cn/simple, Aliyun https://mirrors.aliyun.com/pypi/simple/, USTC https://mirrors.ustc.edu.cn/pypi/simple, Tencent https://mirrors.cloud.tencent.com/pypi/simple, Huawei https://repo.huaweicloud.com/repository/pypi/simple.

One that is easy to miss: uv python install downloads its interpreter from GitHub Releases. Set UV_PYTHON_INSTALL_MIRROR, or just use the system Python — this project wants >=3.12,<3.14, and Ubuntu 24.04 ships 3.12.

Node and npm

Only needed if you build the console or the application image yourself; pulling prebuilt images needs no Node at all.

npm config set registry https://registry.npmmirror.com

npm ci pulls packages with native binaries such as esbuild and rollup. In current versions those ship as platform-specific npm packages, so switching the registry is enough — there is no separate binary mirror to configure.

While on the subject: the Go downloader service needs no GOPROXY. It has no third-party dependencies at all, only the standard library, so its build fetches no modules.

GitHub, and proxies at build time

git clone and CloakBrowser’s pip install "… @ git+https://github.com/…" both reach GitHub. The least intrusive fix is a proxy scoped to GitHub rather than a global one:

git config --global http.https://github.com.proxy http://127.0.0.1:7890

With no proxy to hand, the installer can clone from a mirror instead; read Cloning from a mirror, and what it risks first.

Network access inside a Docker build does not read the host’s git configuration; the proxy has to be passed as a build argument (HTTP_PROXY and HTTPS_PROXY are BuildKit predefined arguments, so the Dockerfile does not declare them):

COMPOSE_ENV_FILES=.env docker compose -p dtk -f docker/compose.yml build \
  --build-arg HTTP_PROXY=http://172.17.0.1:7890 \
  --build-arg HTTPS_PROXY=http://172.17.0.1:7890 \
  browser-rpc

172.17.0.1 is the host as seen from the default bridge network. A common trap here: a proxy listening only on 127.0.0.1 is unreachable from inside the container — it has to listen on 0.0.0.0, or at least on the docker0 address.

Cloning from a mirror, and what it risks

With no proxy to hand, the installer can clone from a mirror of this repository instead. GitHub is always the default; read what switching hands over before you do it:

• The mirror hands you all of the code you install, docker/compose.yml included, and so which images run. Nothing in the installer can tell its copy from GitHub’s: a mirror that is compromised, or run by someone you do not know, can give you other code.
• Every upgrade after this one comes through it too. After the clone the checkout’s origin is the mirror, and dtkctl upgrade keeps fetching from it; nothing switches back to GitHub by itself.
• Use a mirror you run or trust. A self-hosted relay that allows only this repository (a Cloudflare Worker, say) is the clean form; vet any public mirror yourself.

There are two ways to name one. DTK_GH_PROXY is a prefix put in front of the GitHub address, and only an https:// one is taken; DTK_REPO_URL is the whole clone address, and wins when both are set. The script itself has to come through the mirror too, and that download you prefix by hand:

curl -fsSL https://<mirror>/https://raw.githubusercontent.com/Evil0ctal/Douyin_TikTok_Download_API/main/install/install.sh -o install.sh
less install.sh    # all the more for a script that came through a mirror
DTK_GH_PROXY=https://<mirror>/ bash install.sh

Before the clone, the installer says all of this again and asks whether to clone from that mirror; a bare Enter is no, and nothing is changed. Under --yes, having set the variable is the answer.

What it does not reach:

• With the browser container on, its build installs CloakBrowser from GitHub (the pip install above). That does not go through the prefix and still needs the build-time proxy, and a relay that allows only this repository would refuse it anyway.
• The latest release is looked up at api.github.com first. After 10 seconds without an answer the installer reads the tags from the mirror instead, which is only slower.
• Images come from Docker Hub, through the registry mirror set up above.

On an install that already exists, DTK_GH_PROXY does nothing: it fetches from the checkout’s origin. To move it to a mirror or back to GitHub, change that origin (moving to a mirror carries the risks above):

git -C /opt/dtk remote set-url origin https://github.com/Evil0ctal/Douyin_TikTok_Download_API.git

Check it took

Do not go by feel:

docker info | grep -A5 "Registry Mirrors"    # did it read the accelerator
time docker pull redis:8-alpine              # fast or not, this answers it
pip config get global.index-url
npm config get registry

If a step still fails after switching mirrors, run that step alone and read its own error before suspecting the project — everything listed here names the host it could not reach when it fails.

TikTok, once it is running

Everything above gets the install onto the machine. At runtime Douyin is reachable from a mainland exit and TikTok is not, so TikTok needs a proxy - and the proxy has to be on /proxies before the first TikTok identity is minted or imported. An identity’s egress is fixed when it is created: identities made while there was no proxy keep going direct after one is added, and every TikTok request fails with UPSTREAM_UNREACHABLE. The Clash or mihomo inbound used for the build works here too, at the same kind of address. See A platform this server cannot reach directly.

The stack at a glance

Service	Image	Profile	Optional	What it does
postgres	timescale/timescaledb-ha:pg17	default	no	Every durable row: identities, proxies, tasks, archive, request logs, settings
redis	redis:8-alpine	default	no	Task queue, token buckets, rate-limit counters, response cache, console sessions
migrate	dtk-app	default	no	One-shot: applies migrations, exits; everything else waits for it
api	dtk-app	default	no	REST API, MCP, the console, /healthz and /readyz; the only published port
worker	dtk-app	default	no	Claims tasks, drives the scheduler and identity pool, runs background jobs
browser-rpc	dtk-browser-rpc	browser	yes	Headless browser: mints Douyin and TikTok guest identities, optional signing fallback for them
downloader	dtk-downloader	downloader	yes	Go sidecar that writes media files to a volume
transcriber	dtk-transcriber	transcriber	yes	Offline FunASR speech-to-text for stored videos; no network route

Three networks:

• data is declared internal: true. postgres and redis sit on it and have no route off the host at all — not to the internet, not to anything outside the compose project.
• edge is an ordinary bridge network. api, worker, browser-rpc and downloader are on it because they need egress. browser-rpc and downloader are on edge only: neither has any reason to reach the database, so neither can.
• compute is internal: true as well. transcriber is on it and nothing else, so the sidecar can reach neither the internet nor the database; api and worker join it to call the sidecar. Its models are mounted, so it has nothing to fetch.

First install

1. Clone the repository. The compose file, the Dockerfiles and the migrations all live in it, and the default branch is v5. Everything below runs from the repository root:

git clone https://github.com/Evil0ctal/Douyin_TikTok_Download_API.git
cd Douyin_TikTok_Download_API

One file tells you whether you have the right tree — if it is missing you are on the v4 branch and nothing on this page applies:

ls docker/compose.yml

2. Write .env. Hex rather than base64 for the two passwords, because they end up inside connection URLs where + and / would need escaping:

POSTGRES_PASSWORD=$(openssl rand -hex 24)
REDIS_PASSWORD=$(openssl rand -hex 24)
cat > .env <<EOF
DTK_SECRET_KEY=$(openssl rand -base64 48)
POSTGRES_PASSWORD=${POSTGRES_PASSWORD}
REDIS_PASSWORD=${REDIS_PASSWORD}
DTK_DATABASE_URL=postgresql+asyncpg://dtk:${POSTGRES_PASSWORD}@postgres:5432/dtk
DTK_REDIS_URL=redis://:${REDIS_PASSWORD}@redis:6379/0
EOF

.env.example in the repository root lists the bootstrap variables, most of them with a comment (DTK_BACKUP_DIR is missing from it, and a handful of the rest carry no comment); copy from it if you would rather start from a template. Three of its lines — DTK_BIND_HOST, DTK_BIND_PORT and DTK_REDIS_MAXMEMORY — are read by Compose interpolation, so they take effect only if the file is also fed to Compose, as docker/.env, from your shell, or through COMPOSE_ENV_FILES. The third is read by nothing else at all; the first two are application variables as well, but the api container pins its own bind address in environment:, so the copies in this file move the published port through interpolation or not at all. See The two ways Compose reads .env below. CLOAKBROWSER_COMMIT is the same kind of variable, and the file leaves it commented out: docker/compose.yml supplies the value. .env is git-ignored and excluded by .dockerignore, so it reaches the containers as an env file and never as image content.

3. Bring the stack up. The first run builds the dtk-app image, which includes an npm ci and a uv sync, so expect a few minutes:

docker compose -p dtk -f docker/compose.yml up -d --wait

--wait returns only when every service reports healthy. The api healthcheck is a readiness probe, so a successful --wait means the API can actually serve a request.

4. Read the setup token out of the log:

docker compose -p dtk -f docker/compose.yml logs api

The API prints a banner with a URL of the form http://127.0.0.1:8000/setup?token=…. Open it and create the first administrator. The token lives in Redis for 24 hours, is reused rather than replaced across restarts (so a link you are holding keeps working), and is burnt by the fifth failed attempt — four wrong tries are survivable. Once an account exists, the token is deleted and /setup is closed. To get the link again on an instance that still has no account, restart the api container — it reprints the same token. A new token is only issued after the old one expires or is burnt by five failed attempts.

No configuration file needs editing at any point. Everything except the bootstrap variables lives in the database and is edited from the console — see Configuration.

The compose file, service by service

postgres

timescale/timescaledb-ha:pg17, on the data network only. It creates the database dtk owned by the user dtk, with POSTGRES_HOST_AUTH_METHOD=scram-sha-256; the password comes from POSTGRES_PASSWORD in .env. DTK_SECRET_KEY is blanked in its environment: block — the database has no use for the key that encrypts its own contents.

Its data directory is /home/postgres/pgdata/data, not the usual /var/lib/postgresql/data. That is the timescaledb-ha image’s layout; mounting the conventional path instead would leave the real data directory inside the container and lose the database on the first rebuild. If you change the image, check that path first.

Healthcheck: pg_isready -U dtk -d dtk every 10s, 12 retries, 30s start period. stop_grace_period is a full minute so a checkpoint is not cut short.

redis

redis:8-alpine, on the data network only, running as the redis user (the image’s own entrypoint only drops privileges when its first argument is literally redis-server, and this service passes a shell command, so the drop is done in the compose file instead).

It holds the task queue, the scheduler’s token buckets, rate-limit counters, the response cache, the console’s sessions and the setup token. Persistence is appendonly yes plus save 60 1.

Two settings matter:

• --maxmemory ${DTK_REDIS_MAXMEMORY:-320mb}, well under the 512m container limit. The response cache writes one entry per distinct request and keeps it for cache.content_ttl (30 minutes by default); measured entries run 16–200 KB, so a busy instance can put hundreds of megabytes here inside one TTL window. Without a ceiling that ends as an OOM kill of the whole container — queue included — rather than as a cache miss.
• --maxmemory-policy volatile-lru, not allkeys-lru. The queue list carries no TTL and so is never an eviction candidate. What can be evicted is what carries a TTL: cache entries, rate-limit counters, token buckets, sessions — all rebuildable. A queued task is not.

The container refuses to start with an empty REDIS_PASSWORD, and its healthcheck authenticates the same way the application does rather than settling for an unauthenticated PING.

migrate

The same dtk-app image as api and worker, started with the argument migrate. It applies Alembic migrations up to head and exits; restart: "no", and no healthcheck, because a one-shot container’s exit status is its health — which is what service_completed_successfully reads on api and worker.

It exists as its own service rather than as something api and worker each attempt, because concurrent CREATE TABLE races are how a first boot corrupts itself, and because nobody should have to remember to run migrations after a pull. Migrations are idempotent, so it runs on every start.

The DTK_SECRET_KEY gate applies here too: a schema created for a deployment that cannot decrypt its own credentials is worse than a failed migration, and this way the failure arrives before any table exists.

api

The dtk-app image with the argument api, which starts uvicorn dtk.api.app:create_app --factory --no-server-header. It serves:

Path	What it is
/	The React console (a single-page app; unknown paths fall back to index.html)
/api/v1/…	The REST API
/api/setup/…	First-run initialization, closed once an account exists
/mcp/	The MCP endpoint. Keep the trailing slash: /mcp answers 307 Temporary Redirect to it — see MCP and AI agents
/docs	The console’s own API reference page
/swagger, /redoc, /openapi.json	The generated API document, without logging in
/healthz	Liveness. Touches no dependency on purpose
/readyz	Readiness: probes Postgres and Redis, 503 if either is down

It performs no upstream platform requests itself — it authenticates, validates, hands work to the service layer and assembles the reply. That is why it is the only container that publishes a port and the only one that needs to be reachable.

Five variables are pinned in its environment: block and therefore cannot be overridden from .env, because environment: wins over env_file:: DTK_BIND_HOST=0.0.0.0, DTK_BIND_PORT=8000, DTK_CONSOLE_DIR=/app/web/dist, DTK_BACKUP_DIR=/var/lib/dtk/backups and DTK_MEDIA_DIR=/var/lib/dtk/media. Only the first four matter. Nothing reads DTK_MEDIA_DIR — the media path is the constant MEDIA_PATH in dtk.ops.capacity — so the compose file sets it and no code consults it, which is why it appears in none of the variable tables below. The bind address is the container’s own; what limits exposure is the published address, not this value.

Volumes: backup-data read-write at /var/lib/dtk/backups (it lists and serves the archives the worker writes) and media-data read-only at /var/lib/dtk/media. The read-only mount is deliberate: the console has to be able to hand a stored file to a browser, and the API is the only component with authentication, scopes, rate limiting and an audit trail — so it serves the bytes, while the downloader stays a sink and never a relay.

Healthcheck: GET /readyz every 15s, 5 retries, 40s start period. browser-rpc is deliberately not part of readiness — minting is off the request path, so an instance without it is degraded, not unready.

worker

The same image with the argument worker, which runs python -m dtk.worker. It claims tasks from the Redis queue, runs them through the scheduler and the identity pool, and writes results; it also runs the pool filler, the proxy prober, retention, watchlist ticks and archive rechecks.

One worker runs 4 tasks at a time. Real concurrency is bounded by the identity pool, not by that number — extra slots simply wait on the scheduler.

It has no port, so its healthcheck opens a TCP connection to the address in DTK_REDIS_URL: a worker that cannot reach Redis does no work at all, and that is the failure worth detecting.

stop_grace_period is 45s, long enough for an in-flight upstream request to finish and release its identity lease rather than leaving it to expire by TTL. Killing a worker mid-request wastes an identity’s quota.

browser-rpc (profile browser)

A small FastAPI service wrapping a headless browser, with two jobs, both off the hot path: mint Douyin and TikTok guest identities through an identity’s own proxy, and sign their requests with the platform’s own JavaScript when the in-process algorithm has drifted. Its RPC surface is POST /rpc/mint, POST /rpc/sign and GET /rpc/health. Bilibili needs neither job: its guests are minted over plain HTTP by the worker, and its WBI signature is always computed in-process, whatever signing.mode says.

It is the heaviest thing in the stack and the reason the resource limits exist. Warm, it was measured at 2.57 GiB and 629% CPU — six cores’ worth of Chromium. The cap of 4 cores is deliberately below that: signing and minting are off the request path, so this is exactly the service that should yield when the machine is busy.

Everything it writes is RAM:

Mount	Size	Why
/tmp	tmpfs, 3g	The driver passes --disable-dev-shm-usage, so every renderer’s shared memory lands here. Roughly 300 MB per resident context; 1g fitted barely three, and opening a fourth crashed Chromium in a way that reads exactly like a platform block
/profiles	tmpfs, 1g	Single-use mint profiles. A tmpfs makes “never reused, never on disk” structural rather than a promise about cleanup code
shm_size	1gb	Insurance for a future build that stops disabling /dev/shm; it is not the fix for renderer crashes

If you raise DTK_BROWSER_WARM_CONTEXTS, raise the /tmp tmpfs in docker/compose.yml with it: budget roughly 300M × warm_contexts × 2 platforms, plus a mint in flight.

It drops all capabilities and adds back only SYS_ADMIN, which Chromium’s own sandbox needs to create user namespaces. That is the lesser evil against --no-sandbox, which would remove the renderer boundary entirely. init: true because Chromium leaves zombies behind and PID 1 has to reap them.

Its healthcheck reads the status field of the response body, not just the status code: /rpc/health answers 200 even when the browser backend failed to start — deliberately, so you get a reason in the log instead of a restart loop — and a probe that stopped at the status code would call an image built without a browser healthy.

downloader (profile downloader)

A single static Go binary on a scratch base: no shell, no package manager, nothing to execute but the binary. It is the one container that dials arbitrary CDN hosts and writes their bytes to a disk you care about, so it has the least in it. Its healthcheck is the binary probing itself (downloader healthcheck) because a scratch image has no curl, and adding one would undo the reason the image is scratch.

It mounts media-data read-write at /var/lib/dtk/media and is on the edge network only. DTK_SECRET_KEY, DTK_DATABASE_URL and DTK_REDIS_URL are all blanked in its environment: block even though the shared env file carries them: this is the component whose compromise would matter most, and it is the one that needs none of them.

Files land as <platform>/<author>/<content id>/ with a meta.json beside them. Transfers are streamed with io.Copy, so resident memory is a handful of 32 KiB buffers per concurrent file plus the Go runtime — not the size of the files.

transcriber (profile transcriber)

A small FastAPI service around a quantized FunASR pipeline: Paraformer-large for speech, FSMN-VAD to cut it into segments, CT-Transformer for punctuation, then Chinese inverse text normalization. It turns a video the downloader already stored into text; it never fetches media of its own.

Everything about it is shaped by handling files nobody vetted:

• It has no network route. It sits on compute only, which is internal: true. The models are mounted read-only at /models/funasr from the host directory in DTK_TRANSCRIBER_MODEL_PATH, and the container never downloads them — a missing or malformed model is a failed health check, not a silent fetch. Preparing the FunASR ONNX models covers getting them.
• It writes nothing that outlives it. media-data is mounted read-only, the root filesystem is read-only, and ffmpeg streams the audio into the process as samples, so no WAV is left beside your files. The only writable path is a 1 GiB tmpfs on /tmp.
• It holds no secrets. DTK_SECRET_KEY, DTK_DATABASE_URL and DTK_REDIS_URL are blanked, as on the downloader. All capabilities are dropped and it runs as UID 10001.

It runs one job at a time by default (DTK_TRANSCRIBER_WORKERS) with eight queued behind it (DTK_TRANSCRIBER_QUEUE), and refuses with 429 past that; a transcription refused that way is marked failed with the reason, and starting it again later is the fix. Its healthcheck reads the status field of /health, which is degraded while the models are missing or failed to load, so a container that started without them is reported unhealthy instead of accepting jobs it cannot run.

The optional profiles

None of them starts by default. Leaving one off is a supported deployment, not a broken one.

The browser profile gives you automatic Douyin and TikTok identity minting and the browser signing fallback (Bilibili needs neither):

echo 'DTK_BROWSER_RPC_URL=http://browser-rpc:9000' >> .env
docker compose -p dtk -f docker/compose.yml --profile browser up -d --build

Two separate things have to be true: the container has to be running (the profile), and api and worker have to be told where it is (DTK_BROWSER_RPC_URL in .env). Left unset, Douyin and TikTok run manual-import only — you paste cookies in from a browser and those pools work from them, while Bilibili keeps minting its own guests. That is the correct degraded mode rather than an error, and the pool falls back to it on its own when browser-rpc is unreachable. See Identities and proxies.

The downloader profile gives you media stored on your own disk:

echo 'DTK_DOWNLOADER_URL=http://downloader:9100' >> .env
docker compose -p dtk -f docker/compose.yml --profile downloader up -d --build

Same shape: profile plus DTK_DOWNLOADER_URL. Without it, POST /api/v1/downloads answers 501 with a message saying how to turn it on, and nothing else changes — parsing, the archive and the watchlist are all unaffected. Stored media is capped by the media.max_bytes setting, 2 GiB by default; past it the oldest unpinned downloads are removed and an alert says what went. See Downloads, library and watchlist.

The transcriber profile turns stored videos into text on this machine. It needs the downloader, because what it transcribes is a video the downloader saved, and it needs the models on the host first — Preparing the FunASR ONNX models gets and checks them. Then:

cat >> .env <<EOF
DTK_TRANSCRIBER_URL=http://transcriber:9200
DTK_TRANSCRIBER_TOKEN=$(openssl rand -hex 24)
DTK_TRANSCRIBER_MODEL_PATH=/absolute/path/to/funasr-onnx
EOF
COMPOSE_ENV_FILES=.env docker compose -p dtk -f docker/compose.yml \
  --profile downloader --profile transcriber up -d --build

Same shape again, plus the model directory. Without DTK_TRANSCRIBER_URL, POST /api/v1/transcriptions answers 501 and says how to turn it on. COMPOSE_ENV_FILES=.env is not optional here: the model path is read by Compose interpolation, and without it the root .env is never consulted for it — see the two ways Compose reads .env. Give the path as an absolute one: the compose default, ./models/funasr-onnx, is resolved against docker/, not the repository root. With published images, set DTK_TRANSCRIBER_IMAGE=evil0ctal/douyin_tiktok_download_api-transcriber beside DTK_IMAGE and drop --build; a release from before transcription has no such image, so build it or upgrade. If the browser runs on the same host, read When transcription and the browser share a host.

All three at once:

COMPOSE_ENV_FILES=.env docker compose -p dtk -f docker/compose.yml \
  --profile browser --profile downloader --profile transcriber up -d --build

build, config and down have to repeat the profile flags: those commands read the compose file rather than the running stack, so without the flags Compose does not know those services exist — and a down without them stops the core services and leaves the profile containers running. ps and logs <service> find the running containers by project label and need no flags. If you get tired of repeating them, export COMPOSE_PROFILES=browser,downloader,transcriber in your shell.

Building the browser image: the CloakBrowser pin

At least three GitHub organizations publish repositories called cloakbrowser with identical descriptions. “The latest cloakbrowser” therefore does not identify any particular piece of software, so the image takes a repository and a commit, and stamps the pair into DTK_BROWSER_BACKEND_PIN, which /rpc/health reports. “Which browser is this?” is answerable from the running system rather than from build history.

docker/compose.yml supplies both: the CloakHQ repository, and the commit that docker/browser_rpc/requirements.txt — which pins the Playwright that revision needs — was verified with (CloakBrowser 0.6.3, Chromium 146, at the time of writing). They are defaults in the compose file rather than lines in .env, so they move together with those requirements on every git pull, and a Compose build needs nothing set:

docker compose -p dtk -f docker/compose.yml --profile browser build browser-rpc

To build another revision on purpose, set CLOAKBROWSER_COMMIT yourself — a full 40-character commit, never a tag — and CLOAKBROWSER_REPO with it when you build from a fork, because the default commit means nothing against a different repository:

CLOAKBROWSER_REPO=https://github.com/CloakHQ/cloakbrowser \
CLOAKBROWSER_COMMIT="<40-char sha>" \
  docker compose -p dtk -f docker/compose.yml --profile browser build browser-rpc

Then re-check signing and minting: the browser dependencies were verified with the default, not with yours. A value you set stays put while the default moves with the checkout, so take it out again once you no longer need it.

These two are build arguments, resolved by Compose interpolation, so the repository-root .env is the wrong place for them. Put them in your shell as shown above, in docker/.env, or run the build with COMPOSE_ENV_FILES=.env so Compose interpolates from the root file. The next section explains why.

Build docker/Dockerfile.browser with a plain docker build and no --build-arg CLOAKBROWSER_COMMIT=…, and the image still contains the service and the fake backend, but no Chromium: the Dockerfile’s own default is empty, and only Compose supplies the commit. It then refuses to mint unless you set DTK_BROWSER_BACKEND=fake on purpose. Nothing falls back to the fake backend on its own: a pool quietly filling with synthetic identities looks healthy right up to the moment every request comes back risk-controlled.

To pin an upstream image that already carries the browser instead of building from source, pass a digest:

docker build -f docker/Dockerfile.browser \
  --build-arg "BROWSER_BASE_IMAGE=ghcr.io/<org>/cloakbrowser@sha256:<digest>" \
  --build-arg "CLOAKBROWSER_COMMIT=<sha>" -t dtk-browser-rpc:dev .

Build arguments, with the defaults the Dockerfiles declare:

Image	Argument	Default	Meaning
Dockerfile	PYTHON_VERSION	3.12.8 (compose passes the same)	Base Python; bump together with the runtime pin in pyproject.toml
Dockerfile	NODE_VERSION	22	Node used for the console build stage only; it never reaches the runtime image
Dockerfile	UV_VERSION	0.9.18	Pinned, because uv sync --frozen against uv.lock is the reproducibility guarantee
Dockerfile	CONSOLE_OUT_DIR	dist	Checked with test -d, so a Vite config writing elsewhere fails the build instead of shipping an image that serves nothing
Dockerfile, Dockerfile.browser	DTK_UID / DTK_GID	10001	The unprivileged runtime user
Dockerfile.browser	BROWSER_BASE_IMAGE	python:${PYTHON_VERSION}-slim-bookworm	Override with a digest-pinned image that already carries CloakBrowser
Dockerfile.browser	CLOAKBROWSER_REPO	https://github.com/CloakHQ/cloakbrowser	The repository to install from
Dockerfile.browser	CLOAKBROWSER_COMMIT	(empty) (compose passes the commit it names)	The revision. Empty, as on a plain docker build without it, means no browser in the image
Dockerfile.browser	CLOAKBROWSER_INSTALL_CMD	python -m cloakbrowser install	The command that fetches the browser binaries; it belongs to the pinned revision, so verify it when you change the commit
Dockerfile.browser	CLOAKBROWSER_HOME	/opt/cloakbrowser	Where the browser is installed. The rootfs is read-only at runtime, so the entrypoint symlinks it into the container’s HOME
Dockerfile.downloader	VERSION	dev (compose passes ${DTK_IMAGE_TAG:-dev})	Stamped into the binary
Dockerfile.transcriber	PYTHON_VERSION	3.12.8	Base Python for the sidecar; its own dependencies are pinned in docker/transcriber/requirements.txt

Environment variables

The two ways Compose reads .env

This trips up nearly everyone, so it is worth being explicit.

• Service variables come from env_file: ../.env, and Compose resolves that path relative to the compose file. The compose project directory is docker/, so ../.env is the repository-root .env. A missing file is not an error (required: false) — the entrypoint produces the useful message instead.
• Interpolation — the ${…} expressions inside compose.yml itself — does not read that file. It reads docker/.env or your shell. The variables it uses: DTK_BIND_HOST, DTK_BIND_PORT (the published address), DTK_IMAGE, DTK_DOWNLOADER_IMAGE, DTK_TRANSCRIBER_IMAGE and DTK_IMAGE_TAG (which images run), DTK_TRANSCRIBER_MODEL_PATH (the model mount), DTK_REDIS_MAXMEMORY (the redis ceiling) and CLOAKBROWSER_REPO / CLOAKBROWSER_COMMIT (build arguments). The model path is the one that fails quietly: written only into the root .env, it is ignored, Docker mounts an empty docker/models/funasr-onnx in its place, and the transcriber comes up degraded with no models.

So this works:

DTK_BIND_HOST=0.0.0.0 docker compose -p dtk -f docker/compose.yml up -d

and so does pointing Compose at the root file explicitly:

COMPOSE_ENV_FILES=.env docker compose -p dtk -f docker/compose.yml up -d

but putting DTK_BIND_HOST=0.0.0.0 in the root .env and expecting the port to move does not. You can always check what Compose resolved without starting anything:

docker compose -p dtk -f docker/compose.yml config

One more rule on top: environment: beats env_file:. api pins DTK_BIND_HOST and DTK_BIND_PORT to 0.0.0.0:8000 in its own environment: block, so those two values in the root .env change nothing inside the compose stack. They still decide what dtk serve binds to outside Docker.

Bootstrap variables

These are read once at process start, from the environment, because they are needed before the database is reachable. Changing one requires a restart. They are the complete contents of BootstrapSettings in src/dtk/core/config.py, and they all take the DTK_ prefix.

Variable	Default	What it does
DTK_SECRET_KEY	(none — required)	Master key for every stored cookie and proxy credential. At least 32 characters. Generate with openssl rand -base64 48. Changing it makes already-stored credentials undecryptable
DTK_DATABASE_URL	postgresql+asyncpg://dtk:dtk@postgres:5432/dtk	The database. The driver must be asyncpg
DTK_REDIS_URL	redis://redis:6379/0	Queue, cache and counters
DTK_BIND_HOST	127.0.0.1	Listen address. Pinned to 0.0.0.0 inside the api container
DTK_BIND_PORT	8000	Listen port. Pinned to 8000 inside the api container
DTK_LOG_LEVEL	info	Log level
DTK_LOG_JSON	true	Structured JSON logs. Set false for human-readable output while debugging
DTK_BROWSER_RPC_URL	(empty)	Where browser-rpc is. Empty disables automatic Douyin and TikTok minting; those pools then rely on manual imports. Bilibili is minted without it
DTK_BACKUP_DIR	backups	Where backup archives are written and listed from. The image sets it to /var/lib/dtk/backups; the relative default suits a source checkout
DTK_DOWNLOADER_URL	(empty)	Where the downloader sidecar is. Empty disables media downloads entirely
DTK_DOWNLOADER_TOKEN	(empty)	Optional shared secret for that sidecar. Empty means the sidecar checks nothing. It must match on both sides
DTK_TRANSCRIBER_URL	(empty)	Where the transcription sidecar is. Empty disables transcription; POST /api/v1/transcriptions then answers 501
DTK_TRANSCRIBER_TOKEN	(empty)	Shared secret for that sidecar, sent as X-Transcriber-Token. Empty means the sidecar checks nothing. It must match on both sides
DTK_TRANSCRIBER_TIMEOUT_SECONDS	3600	How long the worker waits for one transcription. The sidecar reads the same variable as its ffmpeg decode limit
DTK_TRANSCRIBER_NORMALIZATION_FAILURE_POLICY	fallback	fallback keeps the unnormalized text when number normalization fails; fail fails the transcription instead. The sidecar reads it too

Variables read outside the bootstrap settings

Variable	Default	Read by	What it does
DTK_FORWARDED_ALLOW_IPS	(empty)	docker/entrypoint.sh and dtk.api.routes.support	Comma-separated addresses your reverse proxy connects from. Set it and the entrypoint starts uvicorn with --proxy-headers --forwarded-allow-ips. The application reads it too, to decide whether the per-address login failure limit is enforced — and treats * as if nothing were declared, because an address anyone can forge is a worse lockout key than a shared one
DTK_WORKER_COMMAND	(empty)	docker/entrypoint.sh	Overrides the worker command line entirely. Only needed if the worker module is not importable under either of its usual names
DTK_CONSOLE_DIR	/app/web/dist in the image	dtk.api.console	Where the built console lives. Fallbacks keep a source checkout working without it
DTK_COMMIT, DTK_GIT_COMMIT, GIT_COMMIT	(empty)	dtk.ops.health	Build provenance shown on the console’s system page. Nothing sets them by default

Compose-level variables

Read by Compose interpolation, so they belong in your shell or docker/.env.

Variable	Default	What it does
DTK_BIND_HOST	127.0.0.1	Host address the api port is published on
DTK_BIND_PORT	8000	Host port the api is published on
DTK_IMAGE	dtk-app	Repository for the application image. Set evil0ctal/douyin_tiktok_download_api to pull the published one
DTK_DOWNLOADER_IMAGE	dtk-downloader	Repository for the downloader image; evil0ctal/douyin_tiktok_download_api-downloader when published
DTK_TRANSCRIBER_IMAGE	dtk-transcriber	Repository for the transcriber image; evil0ctal/douyin_tiktok_download_api-transcriber when published
DTK_IMAGE_TAG	dev	Tag for the three images above and for dtk-browser-rpc
DTK_TRANSCRIBER_MODEL_PATH	./models/funasr-onnx	Host directory holding asr/, vad/ and punc/, mounted read-only. The default resolves against docker/, so give an absolute path
DTK_REDIS_MAXMEMORY	320mb	Redis maxmemory. Keep it comfortably under the container’s 512m limit
CLOAKBROWSER_REPO	https://github.com/CloakHQ/cloakbrowser	Build argument for the browser image
CLOAKBROWSER_COMMIT	the revision docker/compose.yml names	Build argument for the browser image. Set it only to choose another revision on purpose: a full 40-character commit

Plus the two the containers require directly, which do live in the root .env:

Variable	What it does
POSTGRES_PASSWORD	Password the postgres container creates the dtk user with. Must match what DTK_DATABASE_URL carries
REDIS_PASSWORD	Password the redis container starts with, and what its healthcheck authenticates with. Must match what DTK_REDIS_URL carries

Getting these out of step with the URLs is a silent failure: the containers start, and only the connection fails. The recipe in “First install” writes both from the same shell variable for exactly that reason.

browser-rpc variables

All prefixed DTK_BROWSER_, all read once at start, all settable in the repository-root .env — the browser container gets the same env_file as the rest of the stack. The three the container fixes for itself (BIND_HOST, BIND_PORT, PROFILE_ROOT) have to agree with expose, the healthcheck and the tmpfs, so they are pinned in environment: and cannot be overridden.

Variable	Default	Meaning
DTK_BROWSER_BACKEND	cloak	cloak or fake. There is no fallback between them
DTK_BROWSER_BACKEND_PIN	set by the image	Repository and commit, reported by /rpc/health
DTK_BROWSER_BIND_HOST	127.0.0.1 (compose pins 0.0.0.0)	Listen address
DTK_BROWSER_BIND_PORT	9000	Listen port. Never published to the host
DTK_BROWSER_PROFILE_ROOT	/tmp/dtk-browser-profiles (compose pins /profiles)	Parent of the single-use mint profiles
DTK_BROWSER_WARM_CONTEXTS	1	Warm signing contexts kept resident, per platform
DTK_BROWSER_WARM_REFRESH_SECONDS	1800	Age at which a warm page is rebuilt; the platform ships new JavaScript regularly
DTK_BROWSER_PREWARM	true	Build warm pages at startup instead of on first use
DTK_BROWSER_MAX_CONCURRENT_MINTS	2	Browsers minting at once (~500 MB each)
DTK_BROWSER_MINT_TIMEOUT_SECONDS	75	Service-side mint budget, below the client’s 90s
DTK_BROWSER_SIGN_TIMEOUT_SECONDS	8	Service-side signing budget, below the caller’s ceiling (signing.rpc_timeout_seconds, default 60)
DTK_BROWSER_CONTEXT_OPEN_TIMEOUT_SECONDS	45	Budget for opening a browser context
DTK_BROWSER_SDK_READY_TIMEOUT_SECONDS	25	How long a fresh page may take to become able to sign, counted after navigation. Too low and you get failures that read like an algorithm change and are really impatience
DTK_BROWSER_SIGN_PROXY_URL	(unset)	Optional exit for the warm signing pages. Without it they load the platform from the host’s own address; minting is unaffected, it always uses the identity’s proxy
DTK_BROWSER_GEO_PROBE_URL	https://ipinfo.io/json	Echo endpoint queried through the proxy to learn the exit country before a context is created. Empty disables it
DTK_BROWSER_GEO_PROBE_TIMEOUT_SECONDS	8	Budget for that probe. It never fails a mint
DTK_BROWSER_DEFAULT_COUNTRY	US	Two-letter code used when neither the caller nor the probe knows
DTK_BROWSER_HEADLESS	true	Off only for debugging on a machine with a display
DTK_BROWSER_LOG_LEVEL	info	Log level for this service

downloader variables

All prefixed DTK_DOWNLOADER_, read once at start by the Go binary, settable in the root .env. BIND and ROOT are pinned by the compose file because they have to agree with expose, the healthcheck and the volume mount.

Variable	Default	Meaning
DTK_DOWNLOADER_BIND	0.0.0.0:9100	Listen address (pinned by compose)
DTK_DOWNLOADER_ROOT	/var/lib/dtk/media	Where files are written (pinned by compose)
DTK_DOWNLOADER_TOKEN	(empty)	Shared secret; must match DTK_DOWNLOADER_TOKEN on the api side
DTK_DOWNLOADER_WORKERS	4	Jobs in parallel
DTK_DOWNLOADER_ITEM_WORKERS	4	Files in parallel within one job. Concurrent transfers are the product of these two
DTK_DOWNLOADER_QUEUE	64	Queue depth
DTK_DOWNLOADER_HISTORY	500	Finished jobs kept in memory for status queries
DTK_DOWNLOADER_MAX_REDIRECTS	5	Redirect ceiling per transfer
DTK_DOWNLOADER_TIMEOUT_SECONDS	900	Ceiling for one transfer

A value that is empty, unparseable or not positive falls back to the default rather than failing the start.

transcriber variables

All prefixed DTK_TRANSCRIBER_, read once at start by the sidecar, settable in the root .env. The two paths are pinned by the compose file because they have to agree with its mounts.

Variable	Default	Meaning
DTK_TRANSCRIBER_MODEL_DIR	/models/funasr	Where the models are read from (pinned by compose; the host side is DTK_TRANSCRIBER_MODEL_PATH)
DTK_TRANSCRIBER_MEDIA_ROOT	/var/lib/dtk/media	Where stored media is read from (pinned by compose)
DTK_TRANSCRIBER_TOKEN	(empty)	Shared secret; must match DTK_TRANSCRIBER_TOKEN on the api side
DTK_TRANSCRIBER_THREADS	2	ONNX Runtime threads per job. Keep it at or under the container’s cpus
DTK_TRANSCRIBER_WORKERS	1	Jobs at once. Each one is a full set of threads
DTK_TRANSCRIBER_QUEUE	8	Jobs waiting behind those; past it the sidecar answers 429
DTK_TRANSCRIBER_JOB_HISTORY	64	Finished jobs kept so a restarted worker can still collect them. Never below workers plus queue
DTK_TRANSCRIBER_PROGRESS_LOG_SECONDS	30	How often a running job logs its progress. At least 5
DTK_TRANSCRIBER_TIMEOUT_SECONDS	3600	ffmpeg decode limit for one file, at least 60. The worker reads the same variable
DTK_TRANSCRIBER_NORMALIZATION_FAILURE_POLICY	fallback	fallback or fail. Any other value leaves the sidecar unhealthy rather than guessing

Runtime settings seeded from the environment

Everything else is a runtime setting: it lives in the settings table, is edited from the console without a restart, and is seeded from the environment once, at first init. The env name is the setting key uppercased with dots replaced by underscores and DTK_ in front, so cache.content_ttl seeds from DTK_CACHE_CONTENT_TTL.

After that first init the environment no longer wins. If you change one of these in .env and nothing happens, that is why — change it in the console instead. The full list is in Configuration.

Volumes

Four named volumes, all managed by Docker. They are named rather than bind-mounted so a default install cannot write into your checkout.

Volume	Mounted at	Written by	What lives in it
postgres-data	/home/postgres/pgdata/data on postgres	postgres	Everything durable: identities, proxies, users, API keys, tasks, the content archive, request logs, settings
redis-data	/data on redis	redis	The append-only file and RDB snapshots: the queue survives a restart, the cache does not need to
backup-data	/var/lib/dtk/backups on api and worker	worker (api reads)	Backup archives. Shared because the worker writes them and the API lists and serves them
media-data	/var/lib/dtk/media on downloader (rw), api and transcriber (ro)	downloader	Downloaded media, laid out as <platform>/<author>/<content id>/ with a meta.json

backup-data and media-data exist because the application containers run with a read-only root filesystem — the only other writable path they have is a 64 MB tmpfs on /tmp. The backup directory is created in the image owned by the runtime user, because Docker seeds a fresh named volume from whatever the image has at that mount point: an absent or root-owned path would yield a root-owned volume a non-root process cannot write, and the error would arrive the moment someone asked for a backup rather than at startup. The downloader image does the same with an empty /seed/media.

To see what they are called and how big they are:

docker volume ls --filter label=com.docker.compose.project=dtk
docker system df -v | grep dtk_

Ports and binding

Only api publishes a port, and by default it publishes to loopback:

ports:
  - "${DTK_BIND_HOST:-127.0.0.1}:${DTK_BIND_PORT:-8000}:8000"

browser-rpc and downloader use expose, which publishes nothing to the host — they are reachable only from other containers on the edge network. postgres and redis publish nothing and sit on an internal network with no route off the host at all. This is the single most important thing the compose file does for your security posture, and it is easy to undo by accident.

Serving the outside world is meant to be an explicit act:

DTK_BIND_HOST=0.0.0.0 docker compose -p dtk -f docker/compose.yml up -d

Do that only with TLS in front of it. The container does not terminate TLS and will not do it for you. If your reverse proxy runs on the same host, leave the default loopback binding and point the proxy at 127.0.0.1:8000; nothing needs to be published on a public interface at all.

Behind a reverse proxy

Two problems appear the moment a proxy sits in front of the API, and both have concrete answers.

The source address. Everything arriving through a TLS terminator carries the terminator’s address, and everything arriving through Docker’s published-port userland proxy carries the bridge gateway’s. Either way request.client.host is a single address shared by the whole internet. DTK_FORWARDED_ALLOW_IPS is how you say otherwise:

# in the repository-root .env, alongside DTK_SECRET_KEY
DTK_FORWARDED_ALLOW_IPS=172.18.0.5

Value	What the api does with the source address
unset (default)	Peer address only. It is still recorded, but no login is refused because of it: the per-address failure counter degrades to an auth.login_spray_suspected warning
your proxy’s address	Real client addresses in the audit trail, the session list and the logs, and the per-address login limit refuses again
*	uvicorn believes X-Forwarded-For from whoever sends it, so the address is forgeable. The api treats this as undeclared and will not refuse a login on it

The per-address limit is conditional on purpose. With one shared address it is one global bucket: twenty failed logins from a stranger would lock every account out of the console for fifteen minutes, repeatably — a denial of service handed to anyone who can reach the login page. The per-account limit is not conditional, because five failures cost an attacker only the account they are guessing at.

The address to use is the one the api container actually sees. If your proxy is a container on a shared network, that is its container address. If it is a process on the host reaching the published port, it is the Docker bridge gateway:

docker network inspect dtk_edge -f '{{ (index .IPAM.Config 0).Gateway }}'

Never set it to * on an instance reachable from the internet.

Streaming. There is no WebSocket anywhere in this project — the console polls every 5–10 seconds, deliberately. There are two streaming responses, and a proxy that buffers will break both:

Endpoint	Content type	Note
GET /api/v1/tasks/{task_id}/events	text/event-stream	Server-sent events until the task settles, up to 300 seconds. A keep-alive comment goes out every 15 seconds of silence. The response already carries Cache-Control: no-cache and X-Accel-Buffering: no
GET /api/v1/archive/export	application/x-ndjson	One post per line, streamed a page at a time so the size of your archive is not what breaks the export

nginx honours X-Accel-Buffering: no, so SSE works with an ordinary proxy block as long as the read timeout exceeds 300 seconds:

server {
    listen 443 ssl;
    server_name dtk.example.com;

    ssl_certificate     /etc/letsencrypt/live/dtk.example.com/fullchain.pem;
    ssl_certificate_key /etc/letsencrypt/live/dtk.example.com/privkey.pem;

    # The API refuses bodies over 1 MiB itself, with a proper error envelope.
    # Keep nginx's limit above that so the caller gets that message.
    client_max_body_size 2m;

    location / {
        proxy_pass http://127.0.0.1:8000;
        proxy_http_version 1.1;
        proxy_set_header Host              $host;
        proxy_set_header X-Real-IP         $remote_addr;
        proxy_set_header X-Forwarded-For   $proxy_add_x_forwarded_for;
        proxy_set_header X-Forwarded-Proto $scheme;
        # Longer than the 300s ceiling on the task event stream.
        proxy_read_timeout 320s;
    }
}

Caddy sets the X-Forwarded-* headers itself and needs only the buffering hint:

dtk.example.com {
    reverse_proxy 127.0.0.1:8000 {
        flush_interval -1
    }
}

The API also sets its own security headers and sends no Server header (--no-server-header). Cross-origin access is closed by default: security.cors_allow_origins is empty, meaning same-origin only. If you serve the console from a different origin than the API you have to open it deliberately — see Security.

Verifying an install

# Every service healthy?
docker compose -p dtk -f docker/compose.yml ps

# Liveness and readiness
curl -s http://127.0.0.1:8000/healthz
curl -s http://127.0.0.1:8000/readyz

# Which revision is on disk, and what the CLI thinks
docker compose -p dtk -f docker/compose.yml exec api dtk --version
docker compose -p dtk -f docker/compose.yml exec api dtk migrate --show

# With the browser profile: is there actually a browser in there?
docker compose -p dtk -f docker/compose.yml exec api \
  python -c "import urllib.request; print(urllib.request.urlopen('http://browser-rpc:9000/rpc/health').read().decode())"

# With the transcriber profile: are the models loaded? "ok", or "degraded" with a reason
docker compose -p dtk -f docker/compose.yml exec api python -c "
import json, os, urllib.request
r = urllib.request.Request('http://transcriber:9200/health',
    headers={'X-Transcriber-Token': os.environ.get('DTK_TRANSCRIBER_TOKEN', '')})
h = json.load(urllib.request.urlopen(r, timeout=5)); print(h['status'], h['detail'] or '')"

/readyz returns 503 with a per-component breakdown when Postgres or Redis is down, which is usually faster to read than the logs. It probes only those two: browser-rpc is deliberately excluded, and so are the other optional containers, for the same reason — none of them is on the request path, so an instance without one is degraded, not unready.

There is also an end-to-end smoke test in the repository. It brings up a separate compose project (dtk-smoke) from the same compose file, walks the initialization flow, mints an API key, exercises the documented contract and tears everything down including volumes:

./scripts/smoke.sh          # full run
./scripts/smoke.sh --keep   # leave the stack up for inspection

Because it uses the same compose file, it publishes the same 127.0.0.1:8000. Stop your main stack first, or the port is already taken.

Upgrading

The least effort is the guided script: ./dtkctl upgrade in its directory. It asks GitHub for the latest release, checks out its tag, pulls it, backs up, migrates and restarts — writing the new version into .env only after the migrations, and putting the checkout and .env back if anything before the restart fails, so the old containers keep running. What follows is every step of that, by hand.

Back up first. The archive is a logical export, not a volume snapshot, and restoring it needs the same DTK_SECRET_KEY — credentials are exported still encrypted.

docker compose -p dtk -f docker/compose.yml exec api dtk backup create -o /var/lib/dtk/backups
docker compose -p dtk -f docker/compose.yml exec api dtk backup list --dir /var/lib/dtk/backups

Pass -o explicitly: the CLI’s own default is a relative backups/ directory, and the container’s root filesystem is read-only, so an unqualified dtk backup create fails inside the container. DTK_BACKUP_DIR is what the console and the worker use, not the CLI. Identities are left out of an archive unless you pass --include-identities, because they are bound to a proxy and an exit address a new machine does not have.

To copy an archive off the volume:

docker compose -p dtk -f docker/compose.yml cp \
  "api:/var/lib/dtk/backups/<file>" ./

Then upgrade:

git pull
export COMPOSE_ENV_FILES=.env    # so compose reads DTK_IMAGE_TAG from .env
docker compose -p dtk -f docker/compose.yml build              # built from source
docker compose -p dtk -f docker/compose.yml pull api worker    # or: on the published images
docker compose -p dtk -f docker/compose.yml up -d --wait

On the published images with DTK_IMAGE_TAG=latest, nothing in .env changes: latest is the newest release. A pinned version or sha- build moves in .env before the pull, and that edit is the price of the pin; it is the one step ./dtkctl upgrade takes off your hands. A checkout the script made sits on a release’s tag rather than on main, so there git pull has no branch to pull; move it with git fetch origin tag v5.1.7 && git checkout --detach v5.1.7 instead, naming the release you are moving to.

migrate runs on every start and Alembic is idempotent, so schema changes apply themselves. Named volumes survive rebuilds.

Operations writes the same upgrade out as a numbered procedure, with the backup first and the /readyz and migrate-log checks afterwards; the points below apply to it too.

Three things to know:

• build and pull only reach the services in the active profiles. To update the browser, downloader or transcriber image, repeat the profile flags - and pull downloader and transcriber by name on the published images, or they keep the old version beside a new API. The browser image is never published, so it is rebuilt on every upgrade, with the CloakBrowser revision the new docker/compose.yml names; when that revision moved, the rebuild downloads a new Chromium and takes several minutes. If your .env still has a CLOAKBROWSER_COMMIT line copied from an older .env.example, delete it unless you chose that revision on purpose, or the export above keeps the build on it while the checkout moves on. Under a tag that does not change (latest, or a source build’s dev) a skipped rebuild leaves the old browser code running. Moving to a tag from before transcription means turning it off first, because that tag has no transcriber image to pull.
• Migrations are forward-only from the CLI. dtk migrate upgrades; there is no downgrade command. Rolling back a release that changed the schema means a new install of the older release with a backup taken before the upgrade restored into it, because a restore fills gaps and does not revert (Operations). ./dtkctl upgrade refuses such a downgrade before it pulls anything.
• Never change DTK_SECRET_KEY as part of an upgrade. Every stored cookie and proxy credential is encrypted with it. If it has to change, retire the affected identities and mint or import them again.

To time an upgrade rather than let it happen at container start, run the migration by hand:

docker compose -p dtk -f docker/compose.yml exec api dtk migrate

Scaling: more workers

docker compose -p dtk -f docker/compose.yml up -d --scale worker=4

Each worker runs 4 tasks concurrently, so four workers is sixteen slots. That is almost certainly not what limits you.

Throughput is bounded by the identity pool, not by the worker count. Every upstream request takes an exclusive lease on an identity and passes a per-(identity, endpoint) token bucket. Workers beyond the number of healthy identities do not add throughput; they queue. Before scaling workers, look at the pool: pool.target_size defaults to 8 per platform, and raising it means minting more identities, which means proxies. See Identities and proxies.

Where extra workers do help: long-running background work (archive rechecks, watchlist ticks, large download jobs) not blocking interactive parsing, and surviving the loss of one worker without a pause. Scale the API instead if you are serving many cached reads — but that means a second published port and a load balancer, which the compose file does not set up for you.

Postgres has no CPU cap and Redis is the smallest thing in the stack, so neither is usually the first ceiling you meet.

Running without Docker

Supported, and how the project is developed. You need:

Requirement	Version
Python	>=3.12,<3.14
uv	Any recent release
PostgreSQL	17 with the timescaledb extension available
Redis	8
Node	22, only if you want to build the console

Clone the repository first; everything below runs from its root:

git clone https://github.com/Evil0ctal/Douyin_TikTok_Download_API.git
cd Douyin_TikTok_Download_API

Bring up a database and a Redis if you do not already have them. The repository ships a test-fixture compose file that publishes them on non-default ports so they cannot collide with anything:

docker compose -p dtk-test -f docker/compose.test.yml up -d --wait
# postgres on 127.0.0.1:55432 (database dtk_test), redis on 127.0.0.1:56379

That fixture stores Postgres data on a tmpfs and is wiped by down -v. It is fine for development and wrong for anything you want to keep.

Install and configure:

uv sync --all-extras

cat > .env <<'EOF'
DTK_SECRET_KEY=replace-me-with-openssl-rand-base64-48
DTK_DATABASE_URL=postgresql+asyncpg://dtk:[email protected]:55432/dtk_test
DTK_REDIS_URL=redis://127.0.0.1:56379/0
DTK_BIND_HOST=127.0.0.1
DTK_BIND_PORT=8000
EOF

Apply the schema, then run the two processes in separate terminals:

uv run alembic upgrade head          # or: uv run dtk migrate
uv run dtk serve --reload            # API on 127.0.0.1:8000
uv run dtk worker                    # the task worker

dtk serve binds to DTK_BIND_HOST/DTK_BIND_PORT unless you pass --host/--port. --reload cannot be combined with --workers, and the command tells you so before it reads the environment.

For the console you have two options. Build it once and let the API serve it:

cd web && npm ci && npm run build      # emits web/dist

The API finds web/dist from a checkout without any configuration; DTK_CONSOLE_DIR overrides the search if your layout differs. Or run the Vite dev server, which proxies /api, /docs, /redoc, /openapi.json, /healthz and /readyz to a locally running API:

cd web && npm run dev                  # http://localhost:5173

DTK_API_TARGET overrides that proxy target if your API is not on 127.0.0.1:8000.

What you give up outside Docker: the read-only root filesystem, the dropped capabilities, the memory and PID ceilings, and the internal network that keeps Postgres and Redis unreachable. Those are properties of the compose file, not of the code. A bare-metal production install has to reproduce them some other way — see Security.

The optional services can still be run by hand. browser-rpc runs from the repository without a container, though only the fake backend works without CloakBrowser installed:

PYTHONPATH=docker DTK_BROWSER_BACKEND=fake DTK_BROWSER_BIND_PORT=19000 \
  DTK_BROWSER_PROFILE_ROOT=/tmp/dtk-browser DTK_BROWSER_GEO_PROBE_URL= \
  uv run python -m browser_rpc

That is useful for exercising the RPC surface, not for producing usable identities — the fake backend produces synthetic cookies no platform accepts, which is exactly why nothing falls back to it on its own.

A production install without Docker

What is above is the development path — it points you at the test fixture, and that fixture keeps its data on tmpfs. To run this on a server you install the database and Redis yourself. The steps below were run end to end on a clean Ubuntu 24.04 host.

1. PostgreSQL and TimescaleDB

TimescaleDB is not in Ubuntu’s repositories, so two of them get added. Without it the first migration refuses to run and says why — the request log, identity events and content snapshots are all hypertables.

sudo apt-get install -y curl ca-certificates gnupg lsb-release

sudo install -d /usr/share/postgresql-common/pgdg
sudo curl -fsSL https://www.postgresql.org/media/keys/ACCC4CF8.asc \
  -o /usr/share/postgresql-common/pgdg/apt.postgresql.org.asc
echo "deb [signed-by=/usr/share/postgresql-common/pgdg/apt.postgresql.org.asc] \
https://apt.postgresql.org/pub/repos/apt $(lsb_release -cs)-pgdg main" \
  | sudo tee /etc/apt/sources.list.d/pgdg.list

curl -fsSL https://packagecloud.io/timescale/timescaledb/gpgkey \
  | sudo gpg --dearmor -o /etc/apt/trusted.gpg.d/timescaledb.gpg
echo "deb https://packagecloud.io/timescale/timescaledb/ubuntu/ $(lsb_release -cs) main" \
  | sudo tee /etc/apt/sources.list.d/timescaledb.list

sudo apt-get update
sudo apt-get install -y postgresql-17 timescaledb-2-postgresql-17 redis-server

2. Actually load TimescaleDB

Skip this and CREATE EXTENSION fails with an error that does not point here:

echo "shared_preload_libraries = 'timescaledb'" \
  | sudo tee -a /etc/postgresql/17/main/postgresql.conf
sudo systemctl restart postgresql

3. Database, user, extension

PGPASS=$(openssl rand -hex 24)
sudo -u postgres psql -c "CREATE USER dtk WITH PASSWORD '${PGPASS}';"
sudo -u postgres createdb -O dtk dtk
sudo -u postgres psql -d dtk -c 'CREATE EXTENSION IF NOT EXISTS timescaledb;'

# Confirm the extension is really there before going further
sudo -u postgres psql -d dtk -tAc \
  "SELECT extname, extversion FROM pg_extension WHERE extname='timescaledb';"

4. The code and its Python dependencies

Run it as a dedicated non-root account — the first thing a bare-metal install has to put back, because the container did it for you.

sudo useradd --system --create-home --home-dir /opt/dtk --shell /usr/sbin/nologin dtk
sudo -u dtk git clone https://github.com/Evil0ctal/Douyin_TikTok_Download_API.git /opt/dtk/app
cd /opt/dtk/app

curl -LsSf https://astral.sh/uv/install.sh | sudo -u dtk sh
sudo -u dtk /opt/dtk/.local/bin/uv sync --frozen --no-dev

--frozen installs exactly what uv.lock pins rather than re-resolving; --no-dev skips the test and lint toolchain.

5. Build the console

The API serves the console, so skipping this leaves you with a working API and a 404 where the front page should be.

sudo apt-get install -y nodejs npm     # Node 22
cd /opt/dtk/app/web && sudo -u dtk npm ci && sudo -u dtk npm run build

From a source checkout the API finds web/dist with no configuration. Point DTK_CONSOLE_DIR at it if your layout differs.

6. Configuration

sudo -u dtk tee /opt/dtk/app/.env >/dev/null <<EOF
DTK_SECRET_KEY=$(openssl rand -base64 48)
DTK_DATABASE_URL=postgresql+asyncpg://dtk:${PGPASS}@127.0.0.1:5432/dtk
DTK_REDIS_URL=redis://127.0.0.1:6379/0
DTK_BIND_HOST=127.0.0.1
DTK_BIND_PORT=8000
DTK_BACKUP_DIR=/opt/dtk/backups
EOF
sudo chmod 600 /opt/dtk/app/.env
sudo -u dtk mkdir -p /opt/dtk/backups

Redis has no password by default. That is defensible while it only listens on loopback, but you are already writing credentials, so set one anyway: add requirepass to /etc/redis/redis.conf and make DTK_REDIS_URL redis://:<password>@127.0.0.1:6379/0.

7. Migrate, and create the first administrator

cd /opt/dtk/app
sudo -u dtk /opt/dtk/.local/bin/uv run dtk migrate

# Password over stdin rather than in argv: argv is visible to every process on the box
printf '%s' 'your-password-here' \
  | sudo -u dtk /opt/dtk/.local/bin/uv run dtk user create admin --role admin --stdin

8. Two systemd services

api and worker are two processes and each needs a unit.

/etc/systemd/system/dtk-api.service

[Unit]
Description=DTK API
After=network.target postgresql.service redis-server.service
Requires=postgresql.service redis-server.service

[Service]
User=dtk
WorkingDirectory=/opt/dtk/app
ExecStart=/opt/dtk/.local/bin/uv run dtk serve
Restart=always
RestartSec=5
# Some of what the container gave you, put back by hand
NoNewPrivileges=true
PrivateTmp=true
ProtectSystem=strict
ProtectHome=true
ReadWritePaths=/opt/dtk/backups

[Install]
WantedBy=multi-user.target

/etc/systemd/system/dtk-worker.service

[Unit]
Description=DTK worker
After=network.target postgresql.service redis-server.service
Requires=postgresql.service redis-server.service

[Service]
User=dtk
WorkingDirectory=/opt/dtk/app
ExecStart=/opt/dtk/.local/bin/uv run dtk worker
Restart=always
RestartSec=5
NoNewPrivileges=true
PrivateTmp=true
ProtectSystem=strict
ProtectHome=true

[Install]
WantedBy=multi-user.target

sudo systemctl daemon-reload
sudo systemctl enable --now dtk-api dtk-worker
curl -s -o /dev/null -w '%{http_code}\n' http://127.0.0.1:8000/healthz    # expect 200

What you gave up, and how to get it back

The container gave you a read-only root filesystem, dropped capabilities, memory and PID ceilings, and an internal network on which Postgres and Redis are simply not reachable from outside. Those are properties of the compose file, not of the code. The units above restore part of it through ProtectSystem and NoNewPrivileges; the rest comes from leaving Postgres and Redis on 127.0.0.1 (the default — do not change it) and leaving DTK_BIND_HOST alone. The full list is in Security.

The optional components

browser-rpc (automatic Douyin and TikTok minting), downloader (media on disk) and transcriber (text from stored videos) are more processes on bare metal, each with its own dependencies: a separate Python environment plus CloakBrowser and Chromium for the first, a Go toolchain for the second, and a third Python environment with ONNX Runtime, ffmpeg and the models for the last. All are optional — without the browser you import Douyin and TikTok cookies by hand, and without the downloader or the transcriber the matching POST answers 501 and says why.

Mixing is entirely fine if you want one of them without installing it by hand: run the application on bare metal and start those sidecars with docker compose --profile browser --profile downloader, then point DTK_BROWSER_RPC_URL and DTK_DOWNLOADER_URL at them. They speak HTTP and do not care where the other side runs. The transcriber mixes the same way with two conditions: it reads videos from the media-data volume, so start it in the same compose project as the downloader that writes them, and it sits on the internal compute network with nothing published, so reaching it from the host means an override that publishes 9200 on 127.0.0.1 only — with a DTK_TRANSCRIBER_TOKEN on both sides, since the port is now reachable from the host.

Running the worker on its own

The worker is the same image as the API with a different first argument, and the same code as uv run dtk worker. It needs three things and nothing else: DTK_SECRET_KEY, DTK_DATABASE_URL and DTK_REDIS_URL. It needs no inbound port and no console build.

Inside the compose stack, run one by hand during an incident:

docker compose -p dtk -f docker/compose.yml run --rm worker

On a second machine, the constraint is not the worker — it is that postgres and redis are on an internal: true network precisely so nothing off the host can reach them. Moving a worker elsewhere means exposing both, with TLS or a private network of your own, and the compose file will not do it for you. Before doing that, re-read the previous section: extra workers rarely help, because the identity pool is the ceiling.

If the image cannot find the worker module it exits with no worker entry point in this image and tells you to set DTK_WORKER_COMMAND to the command line your deployment uses. That escape hatch exists for forks that moved the module; a stock build does not need it.

Starting over

# Stop everything, keep the data
docker compose -p dtk -f docker/compose.yml down --remove-orphans

# Stop everything and delete the volumes: database, queue, backups, media
docker compose -p dtk -f docker/compose.yml down -v --remove-orphans

down -v is irreversible and takes the backup volume with it. Copy any archive you want to keep off the volume first (see “Upgrading”). Add the profile flags if you started the optional services, or their containers are left behind as orphans.

The repository’s Makefile wraps the common ones — make up, make down, make logs, make clean — always with the dtk project name, so nothing stray is left behind.

Where to go next

• Configuration — every runtime setting, what it costs, and when to change it
• Concepts — identities, the scheduler, signing, and why the pool is the ceiling
• Identities and proxies — filling the pool, with or without the browser container
• Users and API keys — after the first administrator
• Operations — backups, retention, logs, alerts and capacity
• CLI reference — every dtk command this page runs, with its full option list
• Security — what the compose file protects you from, and what it does not
• Troubleshooting — when the stack is up and something still refuses to work