udockerd

udockerd is a Docker Engine API-compatible daemon backed by udocker. It's built for running containers on Termux (Android) without any extra privileges.

The daemon runs on the Termux host. The docker CLI (or docker-py, or any Docker Engine API client) can run either directly on Termux, or inside a proot-distro guest, and talks to it over DOCKER_HOST=tcp://127.0.0.1:PORT.

Installation

pip install udockerd

Requires Python 3.12+ and curl/tar executables on PATH (Termux: pkg install curl tar, often preinstalled elsewhere). udocker downloads its own proot/fakechroot execution backends the first time it needs them, however on Termux it is strongly recommended to install the official Termux proot package for best compatibility.

A self-contained Cosmopolitan Libc build is also produced by CI as a single multiplatform executable, requiring no host Python dependency at runtime.

Usage

$ udockerd --help
usage: udockerd [-h] [--host HOST] [--port PORT] [--data-dir DIR] [-v] [-q] [--version]

Docker Engine API-compatible daemon backed by udocker.
Runs proot/fakechroot containers with no root, no namespaces, no cgroups.

options:
  -h, --help       show this help message and exit
  --host HOST      address to bind to (env: UDOCKERD_HOST, default: 127.0.0.1)
  --port PORT      port to listen on (env: UDOCKERD_PORT, default: 2375)
  --data-dir DIR   udocker data directory for images/containers/layers (env: UDOCKER_DIR, default: ~/.udocker)
  -v, --verbose    increase log verbosity (-v: info, -vv: debug, -vvv: debug + proot trace)
  -q, --quiet      suppress all logging except warnings and errors
  --version        show program's version number and exit

examples:
  udockerd
      listen on 127.0.0.1:2375 (default)
  udockerd --host 0.0.0.0 --port 2375
      listen on all interfaces
  udockerd -v --data-dir ~/.udockerd-data
      verbose logging, custom udocker data directory

then point the docker CLI/SDK at it:
  export DOCKER_HOST=tcp://127.0.0.1:2375
  docker ps

With the daemon running and DOCKER_HOST pointed at it, the regular docker CLI and docker-py work as normal:

docker pull alpine
docker run --rm alpine echo hello
docker build -t myimage .
docker exec -it mycontainer sh

docker build only speaks the classic builder protocol, not BuildKit (see Scope below). If your Docker CLI defaults to BuildKit, set DOCKER_BUILDKIT=0.

Scope

udockerd covers the parts of the Docker Engine API used for container and image lifecycle operations, plus docker build. It doesn't aim for full Docker parity.

Implemented: container create/start/stop/rm/list/inspect/logs, exec, attach, TTY sessions, image pull/list/rm/inspect, docker build (classic builder, one flattened layer per build, no build cache), and docker system prune [-a] (containers/images pruned for real; networks/volumes/build-cache report empty).

Not implemented: BuildKit, HEALTHCHECK/ONBUILD/SHELL/--mount=, remote or git build contexts, networks-as-objects, volumes-as-objects, swarm, and compose-level features.

Networking is always effectively --net=host, since proot has no real network namespace and containers share the Termux host's network stack. API responses report this honestly (empty/host IP, no fabricated per-container IPs) rather than faking an isolated network.

How it works

udockerd is a stdlib-only HTTP server (http.server plus ThreadingMixIn) that implements Docker Engine API routes by driving udocker's Python library directly (LocalRepository, DockerIoAPI, and per-container proot/fakechroot engines) rather than shelling out to a udocker CLI. Running containers are tracked in an in-memory registry of pids, pgids, logs, and status. A small compiled C supervisor is prepended in front of every container process to prevent orphaned processes when the daemon exits.

The daemon itself is pure Python, stdlib plus pure-Python udocker, which is what makes it possible to bundle into a single Cosmopolitan Libc executable with no C extensions anywhere in the import graph. Native runtime dependencies like curl, tar, proot, fakechroot, and a C compiler for the supervisor stay external, fetched or invoked on first use rather than bundled in.

Supported platforms

Linux only. proot has no other target, and udockerd fails fast on anything else.

Native x86_64/aarch64 need no extra tooling. The other architectures run the x86_64 build under Blink emulation.

Building from source

pip install -e ".[dev]"   # install with dev deps (pytest, docker SDK, mypy, ruff)

ruff check src tests      # lint
mypy                       # type check (strict mode)
pytest -v                  # run tests (needs a working docker daemon)

Tests build tests/docker/Dockerfile into an unprivileged container, with no host docker socket and no special privileges, mirroring the real Termux deployment. They run udockerd inside it and drive it with the real docker CLI/SDK. If docker isn't available, harness-dependent tests just skip.

The Cosmopolitan-bundled multiplatform executable is built in CI (.github/workflows/build.yml). A Cosmopolitan Python executable gets udockerd (and its udocker dependency) pip-installed and zip-embedded into its Lib/site-packages/, static supervisor binaries for x86_64/aarch64 are built with zig cc and dropped into data/bin/, and chimplink bundles in Blink for the remaining Linux architectures.

Licensing

The code in this repository is licensed under Apache-2.0. udockerd depends on udocker, which is separately licensed. See udocker's repository for its license terms.