Compose Guard

Compose Guard lints Docker Compose files for security holes, misconfigurations and reliability problems before they reach production.

Try it in your browser → The playground runs this exact Python package with Pyodide, so you can lint a Compose file without installing anything. Your YAML never leaves the browser.

Why it exists

Compose files are usually copied from tutorials and grow by trial and error. They reach production with latest image tags, databases published on every network interface, the Docker socket mounted into a container, plain-text passwords in environment:, no healthchecks and no restart policy. docker compose config happily accepts all of that, and nothing complains until a database is found by a scanner or a service stays down after a reboot.

Compose Guard reads your Compose files and tells you, with the exact line and a concrete fix, what to change. It is small, has one dependency (PyYAML), and fits into CI as a gate or as a GitHub code scanning source.

Features

• 15 documented rules for security (privileged mode, Docker socket, dangerous capabilities, disabled seccomp/AppArmor, host namespaces, sensitive host paths, exposed databases, hard-coded secrets) and reliability (unpinned images, healthchecks, restart policies, depends_on conditions, memory limits, fixed container names), plus CD000 for entries Compose itself would reject.
• Accurate line numbers, including through YAML anchors, aliases and << merge keys: a finding inherited from an x- block points at the line where the value is defined.
• Text, JSON and SARIF 2.1.0 output. SARIF findings show up in GitHub code scanning.
• Inline suppressions (# compose-guard: ignore CD005) and a config file with global and per-service ignores and severity overrides. Suppressed findings are counted, never lost.
• CI-friendly exit codes with a configurable threshold (--fail-on).
• Never prints secrets. Findings about environment variables name the key, never the value.
• Safe parsing. YAML is parsed with yaml.compose and SafeLoader; no Python objects are constructed from your file, and alias bombs are rejected.

Installation

Python 3.10 or newer:

pip install compose-guard
compose-guard --version

Or install from a clone:

git clone https://github.com/yagnang-vyas/compose-guard.git
cd compose-guard
python -m venv .venv && . .venv/bin/activate
pip install .
compose-guard --version

Quick start

# Lint compose.yaml (or compose.yml, docker-compose.yaml, docker-compose.yml) in this directory
compose-guard

# Lint specific files, or standard input
compose-guard deploy/compose.yaml deploy/compose.prod.yaml
cat compose.yaml | compose-guard -

# Fail CI on warnings too, and write SARIF for code scanning
compose-guard --fail-on warning --format sarif > compose-guard.sarif

# See every rule with its rationale and fix
compose-guard --list-rules

python -m compose_guard works the same as compose-guard.

Sample output

examples/insecure/compose.yaml is a deliberately insecure small shop stack (web app, PostgreSQL, Redis, a worker and Adminer) with fake credentials. It triggers all 15 rules:

$ compose-guard examples/insecure/compose.yaml
examples/insecure/compose.yaml:6:3 warning CD007 [web] Service has no healthcheck although it publishes ports
    fix: Add a healthcheck that proves the service is serving, for example test: ["CMD-SHELL", "pg_isready -U app"] for PostgreSQL, with interval, timeout and retries.
examples/insecure/compose.yaml:6:3 warning CD008 [web] Service has no restart policy
    fix: Set 'restart: unless-stopped' (or 'always' / 'on-failure').
examples/insecure/compose.yaml:6:3 note CD010 [web] Service has no memory limit
    fix: Set 'mem_limit: 512m' or 'deploy.resources.limits.memory: 512M'.
examples/insecure/compose.yaml:7:5 warning CD001 [web] Image 'example/shop-web:latest' uses the 'latest' tag
    fix: Pin a specific version tag (postgres:16.4) or a digest (image@sha256:...).
examples/insecure/compose.yaml:8:5 note CD015 [web] Service sets a fixed container_name
    fix: Remove 'container_name' and refer to the service by its service name.
examples/insecure/compose.yaml:13:7 error CD006 [web] Environment variable 'DJANGO_SECRET_KEY' has a hard-coded secret value
    fix: Load it from the environment or an untracked .env file (KEY: ${KEY}), or use Docker secrets with a *_FILE variable.
examples/insecure/compose.yaml:15:9 warning CD009 [web] Depends on 'db' only being started, not healthy, although it has a healthcheck
    fix: Use the long form: depends_on: {db: {condition: service_healthy}}.
examples/insecure/compose.yaml:18:3 note CD010 [db] Service has no memory limit
    fix: Set 'mem_limit: 512m' or 'deploy.resources.limits.memory: 512M'.
examples/insecure/compose.yaml:19:5 warning CD001 [db] Image 'postgres' has no tag, so it resolves to 'latest'
    fix: Pin a specific version tag (postgres:16.4) or a digest (image@sha256:...).
examples/insecure/compose.yaml:21:9 error CD005 [db] PostgreSQL container port 5432 is published on all interfaces
    fix: Remove the port mapping and connect over the Compose network, or bind it to loopback, for example '127.0.0.1:5432:5432'.
examples/insecure/compose.yaml:24:7 error CD006 [db] Environment variable 'POSTGRES_PASSWORD' has a hard-coded secret value
    fix: Load it from the environment or an untracked .env file (KEY: ${KEY}), or use Docker secrets with a *_FILE variable.
examples/insecure/compose.yaml:30:5 warning CD008 [db] Restart policy is 'no'
    fix: Set 'restart: unless-stopped' (or 'always' / 'on-failure').
examples/insecure/compose.yaml:32:3 warning CD007 [cache] Service has no healthcheck although other services depend on it (web)
    fix: Add a healthcheck that proves the service is serving, for example test: ["CMD-SHELL", "pg_isready -U app"] for PostgreSQL, with interval, timeout and retries.
examples/insecure/compose.yaml:32:3 note CD010 [cache] Service has no memory limit
    fix: Set 'mem_limit: 512m' or 'deploy.resources.limits.memory: 512M'.
examples/insecure/compose.yaml:35:9 error CD005 [cache] Redis container port 6379 is published on all interfaces
    fix: Remove the port mapping and connect over the Compose network, or bind it to loopback, for example '127.0.0.1:5432:5432'.
examples/insecure/compose.yaml:38:3 note CD010 [worker] Service has no memory limit
    fix: Set 'mem_limit: 512m' or 'deploy.resources.limits.memory: 512M'.
examples/insecure/compose.yaml:40:5 error CD002 [worker] Container runs in privileged mode
    fix: Remove 'privileged: true' and grant only the specific capabilities or devices needed.
examples/insecure/compose.yaml:41:5 warning CD013 [worker] 'pid: host' shares the host's namespace
    fix: Remove the setting so the container gets its own namespace.
examples/insecure/compose.yaml:43:9 error CD011 [worker] Dangerous capability SYS_ADMIN is added
    fix: Remove the capability, or replace it with the narrowest one the service needs.
examples/insecure/compose.yaml:45:9 error CD012 [worker] The seccomp security profile is disabled
    fix: Remove the 'unconfined' option. If a specific call is blocked, use a custom profile that allows only that call.
examples/insecure/compose.yaml:48:9 error CD006 [worker] Environment variable 'PAYMENT_API_TOKEN' has a hard-coded secret value
    fix: Load it from the environment or an untracked .env file (KEY: ${KEY}), or use Docker secrets with a *_FILE variable.
examples/insecure/compose.yaml:50:9 error CD003 [worker] The Docker socket /var/run/docker.sock is mounted into the container
    fix: Remove the socket mount. If the service really needs the Docker API, put a filtering proxy (for example docker-socket-proxy) in front of it and allow only the calls it needs.
examples/insecure/compose.yaml:51:9 error CD014 [worker] Sensitive host path /etc is mounted
    fix: Mount only the specific application directory the service needs, read-only if possible.
examples/insecure/compose.yaml:53:9 warning CD009 [worker] Depends on 'db' only being started, not healthy, although it has a healthcheck
    fix: Use the long form: depends_on: {db: {condition: service_healthy}}.
examples/insecure/compose.yaml:56:3 note CD010 [adminer] Service has no memory limit
    fix: Set 'mem_limit: 512m' or 'deploy.resources.limits.memory: 512M'.
examples/insecure/compose.yaml:57:5 warning CD001 [adminer] Image 'adminer' has no tag, so it resolves to 'latest'
    fix: Pin a specific version tag (postgres:16.4) or a digest (image@sha256:...).
examples/insecure/compose.yaml:58:5 warning CD004 [adminer] Container uses the host's network stack
    fix: Remove 'network_mode: host' and publish only the ports you need with 'ports:'.

10 errors, 11 warnings, 6 notes, 0 suppressed in 1 file

examples/hardened/compose.yaml is the same stack with every finding fixed: pinned images, loopback-only ports, secrets from .env and Docker secrets, healthchecks with condition: service_healthy, restart policies, memory limits, no privileged settings or host mounts.

$ compose-guard examples/hardened/compose.yaml
No problems found in 1 file.

Web playground

Live at compose-guard.netlify.app. web/ holds its source: a landing page and a browser playground that runs this exact Python package with Pyodide (WebAssembly). Your YAML stays in the browser. See web/README.md for the architecture and Netlify deployment.

Rules

Generated from compose-guard --list-rules, which also prints each rule's fix.

Suppressions and configuration

Inline comments

services:
  db:
    image: postgres:16.4
    ports:
      # compose-guard: ignore CD005 -- only reachable through the VPN
      - "10.8.0.1:5432:5432"
    container_name: db  # compose-guard: ignore CD015

# compose-guard: ignore-file CD010

• # compose-guard: ignore ID[,ID...] suppresses those rules on its own line. When the comment is on a line by itself it also covers the line directly below.
• # compose-guard: ignore-file ID[,ID...] anywhere suppresses those rules for the whole file.
• Anything after -- is a free-text reason. Unknown rule IDs are an error, so a typo cannot silently suppress nothing.

Config file

Compose Guard reads .compose-guard.yml from the current directory, or the file given with --config PATH:

ignore: [CD015]          # ignore these rules everywhere
services:
  worker:
    ignore: [CD007]      # ignore these rules for one service
severity:
  CD010: warning         # change a rule's severity
fail_on: warning         # default for --fail-on

Unknown keys or rule IDs are rejected with a list of valid ones. Command-line flags win: --ignore CD010,CD015 adds to the config file's ignore list and --fail-on overrides fail_on.

Suppressed findings always appear in the summary (2 suppressed). --show-suppressed lists them; SARIF output always includes them, marked as suppressed (inSource for comments, external for configuration).

CI usage

As a gate, any CI system only needs the exit code:

compose-guard --fail-on warning

On GitHub, upload SARIF so findings appear as code scanning alerts on the lines they refer to. Code scanning must be enabled for the repository (Settings → Code security), and private repositories need GitHub Advanced Security.

name: Compose Guard
on: [push, pull_request]

permissions:
  contents: read
  security-events: write

jobs:
  compose-guard:
    runs-on: ubuntu-latest
    steps:
      - uses: actions/checkout@v4
      - uses: actions/setup-python@v5
        with:
          python-version: "3.13"
      - run: pip install compose-guard
      - run: compose-guard compose.yaml --format sarif --fail-on never > compose-guard.sarif
      - uses: github/codeql-action/upload-sarif@v3
        with:
          sarif_file: compose-guard.sarif
          category: compose-guard

This repository's own SARIF demo workflow does the same for the insecure example. It only runs when triggered manually, so the example does not create alerts on every push.

JSON output

--format json prints one object with a stable shape (schema_version 1). For this input on standard input:

services:
  db:
    image: postgres:16.4
    restart: unless-stopped
    mem_limit: 1g
    ports:
      - "5432:5432"

$ compose-guard --format json - < compose.yaml
{
  "schema_version": 1,
  "tool": {
    "name": "compose-guard",
    "version": "0.1.2"
  },
  "files": [
    "<stdin>"
  ],
  "summary": {
    "error": 1,
    "warning": 1,
    "note": 0,
    "suppressed": 0
  },
  "findings": [
    {
      "rule_id": "CD007",
      "severity": "warning",
      "file": "<stdin>",
      "line": 2,
      "column": 3,
      "service": "db",
      "message": "Service has no healthcheck although it publishes ports",
      "fix": "Add a healthcheck that proves the service is serving, for example test: [\"CMD-SHELL\", \"pg_isready -U app\"] for PostgreSQL, with interval, timeout and retries."
    },
    {
      "rule_id": "CD005",
      "severity": "error",
      "file": "<stdin>",
      "line": 7,
      "column": 9,
      "service": "db",
      "message": "PostgreSQL container port 5432 is published on all interfaces",
      "fix": "Remove the port mapping and connect over the Compose network, or bind it to loopback, for example '127.0.0.1:5432:5432'."
    }
  ]
}

• findings holds unsuppressed findings, sorted by file, line and rule ID. service is the service name; line and column are 1-based.
• summary counts unsuppressed findings by severity, plus suppressed findings.
• With --show-suppressed, a suppressed array lists suppressed findings with an extra "suppression": "inline" or "config" field.

Exit codes

Limitations

• Each file is linted on its own. Override files are not merged the way docker compose -f a.yaml -f b.yaml merges them.
• ${VAR} interpolation is detected but not resolved, and .env files are not read. Interpolated values are treated as unknown: an interpolated image tag is not called unpinned, an interpolated secret is not called hard-coded, and an interpolated host IP is not judged.
• Compose Guard only sees the Compose file. A healthcheck defined in an image's Dockerfile, or a firewall in front of a published port, is invisible to it; suppress the finding with a reason.
• Suppression comments are found by scanning the raw text, so a # compose-guard: marker inside a multi-line string would also be read as a suppression.
• It does not lint Dockerfiles or Kubernetes manifests and does not fix files automatically.

Prior art

• hadolint lints Dockerfiles, not Compose files.
• Checkov and KICS are broad infrastructure-as-code scanners that include Docker Compose checks among many other formats.
• docker compose config validates and renders a Compose file but does not judge it.

Compose Guard focuses on Compose only and stays small and dependency-light, with accurate line numbers through anchors and merge keys, per-line suppressions and SARIF output.

Development

python -m venv .venv && . .venv/bin/activate
make install   # pip install -e ".[dev]"
make check     # ruff, ruff format --check, mypy --strict, pytest with >=90% branch coverage
make format    # apply formatting and safe lint fixes

• SPEC.md describes the design (especially the loader) and records every judgment call under "Decisions". AGENTS.md holds the ground rules for contributors.
• Each rule has tests/fixtures/bad_cdNNN.yaml and good_cdNNN.yaml. Bad fixtures mark expected findings with # expect: CDNNN and the test checks the exact lines.
• UPDATE_GOLDEN=1 pytest regenerates tests/golden/insecure.json; python scripts/update_readme.py regenerates the sample output and rule table above. A test fails if either is stale.

Releasing

1. Run make check and python scripts/update_readme.py (and python web/scripts/export_snapshot.py if the rule catalogue changed).
2. Update CHANGELOG.md, bump __version__ in src/compose_guard/__init__.py.
3. Tag and publish a GitHub Release (vX.Y.Z). The Publish workflow builds the sdist/wheel and uploads to PyPI via Trusted Publishing.
4. On PyPI, configure a Trusted Publisher for this repository (compose-guard, workflow publish.yml, environment pypi) before the first release upload.

Roadmap

• Merge override files the way docker compose -f does and lint the result.
• Optional .env support for resolving interpolation.
• More rules: read_only root filesystems, no-new-privileges, missing user:, published admin UIs.
• A pre-commit hook and a ready-made GitHub Action.

License

MIT, see LICENSE.