Django Chainsaw

django-chainsaw-mcp

[release] [checks] [MCP tools] [CLI commands] [prompts]

[tests] [coverage] [python] [django] [license]

An MCP server and CLI that analyses a Django project rather than describing it. Not what is in here — what will hurt: what a delete takes with it, which migration breaks the pods still running, which query returns another tenant's row, what a save() sets off three hops away.

There is no model in the loop. Every answer comes from the AST and Django's own app registry, so the same input gives the same output, and nothing leaves the machine. It is an MCP server so an assistant can ask it questions, and a CLI so CI can gate on the answers.

22 checks: 19 need the app registry (16 Django, three of those DRF as well), one needs only Python, one reads FastAPI, one reads SQLAlchemy. Why this exists →

Install

It has to run in an interpreter that can import your project. Everything here reads the app registry, which means django.setup(), your settings and your apps:

/path/to/project/.venv/bin/python -m pip install django-chainsaw-mcp

A plain uvx django-chainsaw-mcp starts and then fails every check, because uvx gives it an isolated environment with no trace of your project. To avoid installing, hand uv the dependencies instead:

uvx --with-requirements requirements.txt --from django-chainsaw-mcp django-chainsaw check

Two environment variables point it at the project:

django-chainsaw project-info proves the setup before anything else, and says which half is missing. Five minutes end to end →

mcp-name: io.github.syrian963/django-chainsaw-mcp

One command

django-chainsaw check --tenant-root myapp.Organisation

44 finding(s): 1 critical, 14 high, 29 medium
CRITICAL--------  [deploy-safety] RemoveField drops 'legacy_code' while code still uses it      shop/0002_remove_product_legacy_code      During a rolling deploy the old pods keep running against the new      schema and will fail.      fix: Ship a release that stops using it, deploy that everywhere,           then ship this migration.

Every analysis, merged, worst first, one exit code.

On a pull request

This is the part that decides whether a tool like this survives. Point tenancy at a five year old project and it returns two hundred candidates; nobody reads two hundred candidates, somebody adds continue-on-error, and it runs forever with nobody looking.

django-chainsaw tenancy --since main     # only what this branch changeddjango-chainsaw tenancy --baseline       # everything old, ratcheteddjango-chainsaw check --sarif out.json   # annotate the diff, on the line

--since compares at the merge base, so a branch that is behind main is not blamed for other people's work. --baseline keeps existing findings in the report and stops them blocking; anything new fails the build, and fixing an old one is reported so the number only ever goes down. Findings are fingerprinted on file plus identity, never the line, so adding an import does not resurrect twenty findings nobody touched.

--sarif writes the format GitHub and GitLab annotate a pull request with, so findings land on the line instead of in a log nobody opens.

baseline.md · cli.md

As an MCP server

claude mcp add django-chainsaw --scope local \  --env DJANGO_CHAINSAW_PROJECT_PATH=/srv/app \  --env DJANGO_CHAINSAW_SETTINGS_MODULE=myproject.settings \  -- /srv/app/.venv/bin/python -m django_chainsaw_mcp.server

Ask it project_info first: the smallest call that proves both the transport and the Django boot. Five prompts carry the ordering the tools do not — before_deploy, why_is_this_slow, what_breaks_if_i_delete, triage, review_this_branch.

Claude Desktop, Cursor, VS Code, Windsurf, Zed, Docker →

Or as one HTML file

django-chainsaw report --out findings.html --title myproject

[The HTML report, grouped by endpoint]

Grouped by endpoint is the view that matters: which pages carry this, and through what call path. No server, no network, no build step — the CSS, the script and the data are all in the file, so it works from a CI artifact or an email attachment. report.md

The checks

36 of the 37 tools declare readOnlyHint, so a client can stop asking permission for each call; the exception is api_contract_check with update=True, which writes the snapshot and says so. Every tool, argument and output shape →

What it will not tell you

Nothing here executes the target project or reads its data, which buys safety and speed and costs certainty. Every tool states its own blind spots in its own output:

• delete_impact does not run signals or custom delete() overrides.
• find_n_plus_one reports candidates; it reads the template and the model graph, not the queryset in the view.
• migration_risk does not know row counts, PostgreSQL version, or deploy strategy.
• deploy_safety cannot see getattr(obj, name), runtime SQL, or another repository. CLEAR means nothing was found here.

A confident wrong answer is worse than an incomplete one. In this kind of tooling the failure mode is not a crash, it is a plausible sentence that sends someone in the wrong direction. limitations.md

What it does to your code

It imports the target project. django.setup() imports your settings and every app in INSTALLED_APPS, and the checks additionally import the modules that declare serializers, views and URLs — so anything those do at import time happens. Do not point this at code you would not run.

It does not run your application: no view, no task, no management command. One check reads the database, read-only — MigrationLoader reads django_migrations, and nothing is written. It writes files only when you ask: fix --write applies the mechanical class of fix only, and a baseline, a contract snapshot or --sarif write where you tell them to. Nothing leaves the machine — no network calls, no telemetry, no uploads.

SECURITY.md

Documentation

docs/ is the index.

Contributing

A proposal for a new check answers four questions, which the issue template asks directly: what the defect looks like as code, how it fails in production, what already finds it, and what it must stay silent on. Two finished features were deleted from this repository after measurement showed they could not tell a real finding from a correct one.

MIT. One process stays bound to the first project it loads, because django.setup() cannot be undone — run a second instance for a second project.