stackdoctor

Claude finds that the worker shut down and 40 jobs are waiting, in one diagnose() call.

Strictly read-only. stackdoctor never writes, deletes, restarts, retries, revokes or sends anything. Postgres sessions are forced read-only by the server, Redis commands go through an allowlist, and Celery is only inspected. Secrets are redacted from every response.

stackdoctor is an MCP server that lets Claude, Cursor or any MCP client diagnose a Python backend stack (Postgres, Celery, Redis and logs) from a single tool call.

Ask "why are my jobs stuck?" and the assistant calls diagnose(symptom). It runs every relevant check in parallel and returns one timestamped snapshot with:

• findings: blocked queries, missing workers, queue backlogs, log error spikes, Redis memory pressure, …
• a merged timeline of events from Postgres, Celery, Redis and your logs
• possible cause → effect chains when events line up in time, each with its evidence timestamps

possible cause (confidence: medium): Database lock → blocked queries → stuck or failing tasks
  cause:  21:58:01 postgres.lock_held      pid 7959 holds a lock blocking 2 sessions: LOCK TABLE orders …
  effect: 21:58:02 logs.db_lock_wait       [postgres] process 7962 still waiting for RowExclusiveLock …
  effect: 21:58:21 celery.task_failed      tasks.process_order[…] failed: LockNotAvailable: lock timeout
  effect: 21:58:21 celery.task_long_running  tasks.process_order[…] running 13.2s on worker1
  effect: 21:58:34 celery.workers_saturated  worker1: all 2 slots busy, 2 reserved
  caveat: Inferred from timing only. Verify before acting; this is not a confirmed root cause.

Install

You need uv:

curl -LsSf https://astral.sh/uv/install.sh | sh                          # macOS / Linux
powershell -c "irm https://astral.sh/uv/install.ps1 | iex"                # Windows

Then one command runs the server, with no other install step:

uvx stackdoctor

Until the package is on PyPI, run it from a checkout with uvx --from /path/to/stackdoctor stackdoctor, or from git with uvx --from git+https://github.com/lepri89/stackdoctor stackdoctor.

macOS 12 (Monterey): uvx stackdoctor fails with realpath: command not found, because uv's launcher script needs realpath, which only ships with macOS 13+. Use uvx --from stackdoctor python -m stackdoctor instead. In client configs that means "args": ["--from", "stackdoctor", "python", "-m", "stackdoctor"]. Alternatively, run uv tool install stackdoctor once and use stackdoctor as the command.

Configure

Set environment variables in your MCP client config, or in a .env file. stackdoctor looks for .env in the working directory, or uses the file named by STACKDOCTOR_ENV_FILE. Any source you don't configure is skipped, so it's fine to start with only DATABASE_URL.

LOG_SOURCES entries are file paths, docker:<container>, or a bare container name. Only configured sources can be read.

Celery workers: point LOG_SOURCES at the worker's stdout. Celery prints worker: Warm shutdown (MainProcess) straight to stdout, not through logging, so it never reaches a --logfile. Docker containers already capture stdout. For a file, redirect stdout to it (celery … worker >> worker.log 2>&1, or supervisor/systemd stdout capture) so stackdoctor can see when a worker stopped.

Claude Desktop

~/Library/Application Support/Claude/claude_desktop_config.json (macOS) or %APPDATA%\Claude\claude_desktop_config.json (Windows):

{
  "mcpServers": {
    "stackdoctor": {
      "command": "uvx",
      "args": ["stackdoctor"],
      "env": {
        "DATABASE_URL": "postgresql://stackdoctor_ro:password@localhost:5432/app",
        "REDIS_URL": "redis://localhost:6379/0",
        "CELERY_BROKER_URL": "redis://localhost:6379/0",
        "CELERY_RESULT_BACKEND": "redis://localhost:6379/1",
        "LOG_SOURCES": "docker:api,docker:worker"
      }
    }
  }
}

GUI apps often don't see your shell PATH. If Claude Desktop can't find uvx, use the full path: /Users/<username>/.local/bin/uvx on macOS, or C:\\Users\\<username>\\.local\\bin\\uvx.exe on Windows.

Claude Code

claude mcp add stackdoctor \
  -e DATABASE_URL=postgresql://stackdoctor_ro:password@localhost:5432/app \
  -e REDIS_URL=redis://localhost:6379/0 \
  -e CELERY_BROKER_URL=redis://localhost:6379/0 \
  -e LOG_SOURCES=docker:api,docker:worker \
  -- uvx stackdoctor

Or, from a project directory with a .env file, simply claude mcp add stackdoctor -- uvx stackdoctor.

Cursor

.cursor/mcp.json in your project, or ~/.cursor/mcp.json globally. This uses the same mcpServers format as Claude Desktop:

{
  "mcpServers": {
    "stackdoctor": {
      "command": "uvx",
      "args": ["stackdoctor"],
      "env": { "STACKDOCTOR_ENV_FILE": "${workspaceFolder}/.env" }
    }
  }
}

VS Code (Copilot agent mode)

.vscode/mcp.json:

{
  "servers": {
    "stackdoctor": {
      "type": "stdio",
      "command": "uvx",
      "args": ["stackdoctor"],
      "env": { "STACKDOCTOR_ENV_FILE": "${workspaceFolder}/.env" }
    }
  }
}

Tools

Postgres checks are intentionally light. For deep Postgres health (index advice, vacuum, bloat), run Postgres MCP Pro next to stackdoctor.

Safety

stackdoctor is built so that a confused or prompt-injected assistant still can't change your systems.

Postgres

• Every session is opened with default_transaction_read_only=on, statement_timeout=5s and lock_timeout=2s, passed as connection options so they override your URL. stackdoctor also checks the setting and refuses to run if the session isn't read-only.
• run_select parses SQL with sqlglot, not regexes. It accepts exactly one SELECT / WITH … SELECT / EXPLAIN statement and rejects:
  • writable CTEs, SELECT … INTO and FOR UPDATE/SHARE
  • EXPLAIN ANALYZE, because it executes the query
  • functions with side effects, even inside a read-only transaction: pg_terminate_backend, pg_cancel_backend, pg_reload_conf, pg_read_file, pg_read_binary_file, pg_ls_dir, lo_*, dblink*, set_config, pg_advisory_*, query_to_xml (which runs SQL from a string), pg_sleep, …
• Results are always limited: queries are wrapped as SELECT * FROM (<q>) AS sd_sub LIMIT n (default 100, max 1000).

Recommended: a dedicated read-only role. Defense in depth, so the database enforces the rules too:

CREATE ROLE stackdoctor_ro LOGIN PASSWORD 'change-me';
GRANT CONNECT ON DATABASE app TO stackdoctor_ro;
GRANT USAGE ON SCHEMA public TO stackdoctor_ro;
GRANT SELECT ON ALL TABLES IN SCHEMA public TO stackdoctor_ro;
ALTER DEFAULT PRIVILEGES IN SCHEMA public GRANT SELECT ON TABLES TO stackdoctor_ro;
GRANT pg_monitor TO stackdoctor_ro;          -- see other sessions' queries and pg_stat_statements
ALTER ROLE stackdoctor_ro SET default_transaction_read_only = on;

Anything the role can SELECT, the assistant can read. If some tables hold data you don't want to share with an AI, leave them out of the GRANT SELECT.

Redis. Every command is checked against an allowlist that is aware of subcommands (INFO, SCAN, TYPE, PTTL, LLEN, GET, …). CLIENT LIST is allowed; CLIENT KILL / PAUSE are not. The same goes for OBJECT, MEMORY and CONFIG (only CONFIG GET). KEYS is never used; listing always goes through SCAN.

Celery. Only inspect (ping, active, reserved, active_queues, stats, query_task) and broker reads are used: LLEN on Redis, and passive queue_declare on RabbitMQ.

RabbitMQ support is experimental. Queue lengths via passive queue_declare and the inspect calls are implemented but not yet covered by CI. The Redis broker is the tested path. Reports welcome. stackdoctor never revokes, retries, shuts down or sends tasks. To be precise: inspect works by publishing a broadcast message on Celery's control channel (celery.pidbox) and reading replies from a temporary reply queue. That is how Celery's own celery inspect command works. It doesn't touch your task queues or results, but it isn't zero traffic. Everything else stackdoctor does is pure reads.

Redaction. Every response is redacted before it leaves the server. This covers:

• passwords in URLs (postgres://user:***@host)
• Authorization headers and bearer tokens
• PASSWORD=, SECRET_KEY=, api_key: and other secret-looking keys
• AWS, GitHub, Slack, Stripe, OpenAI/Anthropic-style and Google keys, JWTs, and private keys

Task args/kwargs and Redis value previews are truncated and redacted too. Redaction is pattern-based, so treat it as a safety net, not a guarantee.

Output caps. Every tool caps its response (MAX_OUTPUT_CHARS) by trimming lists and long strings, so a huge table or log can't flood the assistant's context.

Logs. Only sources listed in LOG_SOURCES can be read, so the assistant can't ask for /etc/passwd. Docker logs are read with docker logs, called without a shell.

Try the demo

The demo is a small FastAPI + Celery + Redis + Postgres app with scripts that break it in realistic ways. CI runs every scenario on each push (demo.yml): it breaks the stack, calls diagnose() through a real MCP stdio client and checks the findings and chains. To try it without Docker (local Postgres, Redis and worker), see DEMO.md.

cd demo
docker compose up -d --build        # first run builds the app and loads 1M rows (~1 min)
cp .env.example .env                # stackdoctor config for the demo (uses the read-only role)

Point your MCP client at the demo by setting STACKDOCTOR_ENV_FILE to the absolute path of demo/.env. Then break something and ask:

Colima (macOS). docker compose works unchanged once Colima is running: colima start --cpu 2 --memory 4. On macOS 13+ you can add --vm-type vz. On macOS 12 Colima uses QEMU, which must be installed first (brew install qemu). Windows: use Docker Desktop or Docker in WSL2 and run the .ps1 scripts from PowerShell.

Why stackdoctor instead of separate Postgres, Celery and log MCPs?

Development

uv sync
uv run pytest                                   # unit tests (safety layer, diagnose, logs)
STACKDOCTOR_TEST_DATABASE_URL=postgresql://shop:shop@localhost:55432/shop \
STACKDOCTOR_TEST_REDIS_URL=redis://localhost:56379/0 \
uv run pytest                                   # plus live tests against the demo

Releasing

Releases go to PyPI from GitHub Actions with trusted publishing, so no API token is stored anywhere.

1. One-time: on PyPI, add a pending publisher (Account → Publishing) with project stackdoctor, owner lepri89, repository stackdoctor, workflow release.yml and environment pypi. In GitHub, create an environment named pypi (Settings → Environments); adding yourself as a required reviewer is a good idea.
2. Bump version in pyproject.toml, commit, then tag and push: git tag v0.1.0 && git push origin v0.1.0. The workflow checks that the tag matches the version, runs the tests, builds, and publishes.

License

MIT