Invigil MCP Server

Invigil

A CI quality gate that grades a repo against a product-quality doctrine — not code style.

Linters check your code. Dependabot checks your dependencies. Nothing checks whether your project is legible — whether someone arriving cold can act on it: boot it in ten minutes, get an error that tells them the fix, install the thing on PyPI today, read a README that's still a landing page and not a 600-line wall.

That is the test every open-source project takes when someone new finds it — a new engineer, or increasingly an AI agent with a context window instead of patience. If they can't get to "hello world" in 10 minutes, they leave. If the artifact on PyPI is broken because CI only tests the source tree, they leave. If the error message is a silent stack trace, they leave.

Invigil turns those promises into mechanical, exact-fix-reporting checks and runs them in CI — so the project speaks for itself.

invigilate — to watch over an exam and enforce its rules.

---

How Invigil fits in your stack

Invigil does not replace your existing tools; it covers the product-quality gaps they leave behind.

---

Why

You already wrote the doctrine; you just enforce it by hand. Every failing Invigil check tells you what's wrong, why it matters, and the exact command to fix it — because a gate that can't tell you how to pass it is the same broken-error-message anti-pattern it's meant to catch.

It grades against seven Gates, each a legibility promise to a different cold-start reader:

A repo reaches Gn only when every mandatory check for gates ≤ n passes, and gets a letter grade from its weighted score.

Install

One tool, four doors — pick the one that matches where you run it:

Every release ships all of it signed: cosign-signed wheel, sdist, and container image, plus an SPDX SBOM — verifiable with cosign verify against the GitHub OIDC identity.

Quick Start

Run it locally on any repo:

pip install invigil
invigil score .            # human-readable scorecard + the exact fix for every failure
invigil score . --format markdown   # a PR-comment-ready table
invigil score . --format json       # machine-readable

Add it to CI as a report-only gate (posts a scorecard comment + badge, never blocks a PR):

# .github/workflows/quality-gate.yml
name: Quality gate
on: [pull_request]
permissions: { contents: read, pull-requests: write }
jobs:
  invigil:
    runs-on: ubuntu-latest
    steps:
      - uses: actions/checkout@v4
      - uses: invigil/invigil@v1        # the doctrine scorecard
        with:
          enforce: "false"              # flip to true once the grade is stable

Flip enforce: "true" (or set project.enforce: true in .invigil.yml) when you're ready for it to block merges below the target gate.

The action exposes the rendered report and badge as step outputs — pipe the scorecard into the job summary or publish the badge JSON wherever shields.io can reach it:

      - uses: invigil/invigil@v1
        id: invigil
        with: { comment: "false" }
      - run: cat "${{ steps.invigil.outputs.report }}" >> "$GITHUB_STEP_SUMMARY"
      # steps.invigil.outputs.badge → path to the shields.io endpoint JSON

How it works

Two layers, matching the doctrine:

• Static Doctrine Scorecard (every PR, seconds) — inspects the repo filesystem and GitHub metadata: LICENSE, README length, .env.example, a deep-health endpoint, a global error-id handler, SHA-pinned actions, an enforced lockfile, a coverage gate, a daily published-artifact smoke test, ≥5 good-first-issues, docs index, llms.txt/AGENTS.md, and more. Emits text / JSON / Markdown / a shields.io badge.
• Cold-Start Gate (nightly, reusable — invigil stranger) — on a clean runner, installs and boots each published artifact you declare and probes its core surface within a 10-minute budget. Web services get HTTP probes; a CLI image (an artifact with a command:) is run to completion and must exit 0. One reusable workflow replaces the 60-line smoke-published.yml every repo copy-pastes:

# .github/workflows/stranger-gate.yml
name: Cold-start gate
on:
  schedule: [{ cron: "0 3 * * *" }]
  workflow_dispatch:
jobs:
  stranger:
    uses: invigil/invigil/.github/workflows/stranger-gate.yml@v1

Fix by PR (Dependabot-for-legibility)

Opt in to a scheduled bot that applies Invigil's mechanical fixes on a work branch and opens one batched PR — governance scaffolds, agent context files, config hygiene. Three anti-noise rules are built in: it's opt-in only, one stable branch means one PR (never five), and a PR you close unmerged is a "no" the bot respects — it stays silent until you delete the invigil/fixes branch.

# .github/workflows/legibility-fixes.yml
name: Legibility fixes
on:
  schedule: [{ cron: "0 6 1 * *" }]   # monthly — these are one-time scaffolds, not deps
  workflow_dispatch:
jobs:
  fix:
    uses: invigil/invigil/.github/workflows/fix-pr.yml@v1

Under the hood it runs invigil score --fix --pr-mode: the fix engine's CI-lockout stays in force for protected branches — --pr-mode only permits fixes on a non-default branch, so nothing automated ever lands on main without a human merging the PR.

Configuration

Drop a .invigil.yml at the repo root. It's optional for the static scorecard (sensible defaults apply) and required for the Cold-Start Gate (it declares what to boot and probe). Full schema in schema/invigil.schema.json; examples in examples/.

version: 1
project:
  name: my-app
  language: python
  min_gate: G4
  enforce: false
artifacts:
  - type: pypi
    name: "my-app[all]"
  - type: ghcr
    image: ghcr.io/me/my-app:latest
    port: 8000
probes:
  - { url: "/", expect_status: 200 }
  - { url: "/api/things", expect_json_count: { min: 5 } }
boot_budget_minutes: 10

Lightweight & modular

A gate developers bypass is dead weight, so Invigil is built for zero friction:

• Fast offline groups for pre-commit — each check is tagged local/network/heavy. invigil check layout runs the filesystem checks with no network in ~120ms:

  # .pre-commit-config.yaml
  - repo: https://github.com/invigil/invigil
    rev: v1        # tracks the latest v1.x.y
    hooks: [{ id: invigil-layout }, { id: invigil-secrets }]
  

  Heavier, network-bound checks (scorecard, the Cold-Start Gate) stay in CI. invigil score --offline / --layer local / --group supply-chain slice it any way.

• Profiles, so it bends instead of forking. profile: strict | progressive | light, plus per-check weights, optional (ding without gating), and thresholds.fail_on. Make it your doctrine, not a hardcoded one.

• Resilient by design. A scorecard.dev timeout is a SKIP that's excluded from the grade — never a false A-to-C downgrade, never a crashed build.

• AI-era native. The ai group checks that your llms.txt/AGENTS.md leak no secrets and that agent code declares its tool inventory — the first slice of "what's the blast radius if this agent is prompt-injected?"

When your user is an agent

Legibility now has two audiences. The reader arriving cold at your repo is, more often than not, an AI agent: it has a context window instead of patience, exit codes instead of intuition, and it acts only on what the repo states machine-readably. The ai check group grades that surface — not just presence of llms.txt/AGENTS.md, but whether an agent can actually act on them:

Two artifacts fall out of it: an ai-ready badge (shields endpoint, emitted next to the grade badge by --badges-dir) and invigil score --format llm — a deterministic report under ~1 KB, built to be read by an agent: a healthy repo costs it two lines of context.

MCP server

Agents don't have to shell out — Invigil speaks MCP natively (mcp-name: io.github.invigil/invigil):

pip install "invigil[mcp]"    # optional extra; the core CLI stays zero-dep
invigil mcp                   # stdio server

// e.g. .mcp.json for Claude Code — any MCP client works
{ "mcpServers": { "invigil": { "command": "uvx", "args": ["--from", "invigil[mcp]", "invigil", "mcp"] } } }

Three read-only tools: evaluate_repo (the scorecard, llm or json format), check_group (one fast offline group), and preview_fixes (the mutation plan --fix would apply — the agent applies changes with its own edit tools, so nothing here writes).

The doctrine

Invigil encodes a specific product-quality doctrine (the Silent User Doctrine and its Five Disciplines): absence of complaints is not absence of problems — silence is the loudest negative signal a project gets. You test at release time; users arrive after dependencies drift and registries change. Only automation is awake then. Invigil is that automation.

Contributing

Issues and PRs welcome — see CONTRIBUTING.md and good first issues. Invigil grades itself in CI (self-score job); a PR that lowers Invigil's own grade won't merge.

License

Apache-2.0 — see LICENSE.