Skip to main content
Glama

Linters check that code compiles. PrufGate looks for code that compiles, passes review at a glance and still does not do the job. This is the kind of code an AI coding agent (or a hurried human) writes to make a task look finished:

  • a catch block that returns a hard-coded success object;

  • a click handler that does nothing;

  • a test that asserts true, or mocks the very function it tests;

  • a .skip, @ts-ignore or lowered coverage threshold slipped into a diff;

  • an import of a package that is not declared anywhere;

  • an API route that loads a record by ID without checking who owns it.

PrufGate is deterministic: plain Python, no network calls and no LLM unless you ask for one. It runs in seconds on a pull request, and writes SARIF for code scanning and Markdown for humans. It is meant to sit next to your linters and SAST tools, not replace them.

The six engines

#

Engine

What it catches

1

Anti-fabrication (prufgate/anti_fabrication)

Canned success on error paths, simulated return values, MOCK_*/SAMPLE_* data in production code, stub UI handlers, fake "fresh" timestamps, unimplemented stubs, LLM prompts that ask for a verdict without passing any evidence, documentation addresses in production.

2

Test integrity (prufgate/anti_reward_hacking)

Tests with no assertions, assertions that cannot fail, tests that mock the unit under test, swallowed exceptions, assertions deleted or weakened in a diff, test files deleted or test cases removed in a diff, production code special-casing a test's literal input.

3

Diff guard (prufgate/diff_guard)

Skips and focus (.skip, .only, xit, pytest.mark.skip, t.Skip(), #[ignore], @Disabled) added without an issue reference, fake exits (sys.exit(0), process.exit(0), module-level SkipTest) added to tests or test setup, bare suppressions (@ts-ignore, # noqa, // nolint, #[allow]), as unknown as T casts, weakened linter/CI/coverage settings, and new exemptions in PrufGate's own config.

4

Contract guard (prufgate/contract_guard)

Imports not declared in package.json, pyproject.toml/requirements*.txt, go.mod or Cargo.toml; environment variables read without a default and missing from .env.example; frontend calls to API routes that do not exist.

5

BOLA / IDOR (prufgate/bola)

Route handlers that look up a record by a client-supplied ID with no auth guard or no tenant/owner scoping, across Express, Next.js, NestJS, FastAPI, Flask, Django, Go (net/http, Gin, Echo, Chi), Spring, Rails and Laravel. Optional live two-token probing of a running service you own.

6

Diff reviewer (prufgate/review)

High-precision rules on changed lines (SQL built from strings, shell injection, disabled TLS verification, hard-coded credentials, unsanitised HTML, ignored Go errors, credential logging), GitHub suggestion blocks, and optional delegation to an AI CLI or API you choose.

Details, rule IDs, supported languages and known limits for each engine: docs/ENGINES.md.

PrufGate uses pattern and structure heuristics, not a full parser for every language. It will miss things and it will sometimes flag correct code. Every finding names its rule, and inline prufgate-ignore: rule -- reason comments, exemptions and baselines let you record why.

Related MCP server: system-brain-mcp

Quick start

PrufGate needs Python 3.9 or newer and Git.

The documentation home covers this checkout. For unreleased options, install/build a reviewed checkout and keep CLI, Action and image implementations aligned; see installation and upgrading. Published 1.4.0 artifacts retain that release's behavior.

pipx install prufgate     # or: pip install prufgate  (the `prufgate` and `prufgate-mcp` commands)
prufgate --self-check     # every engine against its built-in fixtures
uvx prufgate --version    # or run it without installing, with uv

Or download one file, prufgate.pyz, from the latest release (releases after 1.0.0). It is a zipapp with PyYAML (pure Python) inside, so it needs nothing but Python 3.9 or newer:

curl -LO https://github.com/GuardianVigil-Labs/Prufgate/releases/latest/download/prufgate.pyz
curl -LO https://github.com/GuardianVigil-Labs/Prufgate/releases/latest/download/SHA256SUMS
sha256sum -c --ignore-missing SHA256SUMS
gh attestation verify prufgate.pyz --repo GuardianVigil-Labs/Prufgate
python3 prufgate.pyz --self-check
python3 prufgate.pyz check .

The hooks that prufgate init --hooks writes call prufgate on PATH, so for them install the file as a command: install -m 755 prufgate.pyz ~/.local/bin/prufgate. The zipapp runs every command a pip install has except mcp (use pipx install prufgate for the MCP server). Like a pip install, it does not include the shell batteries (./prufgate.sh fast / full with ast-grep, gitleaks and zizmor); prufgate.pyz fast and full run the Python engines only. For the full battery use the container image or a checkout. There are no native per-OS binaries: they would need a build on three operating systems for every release, and the zipapp covers the same need. python3 scripts/build_pyz.py builds the same file from a checkout; it downloads the hash-locked PyYAML sdist from PyPI and refuses one that does not match requirements.txt.

Or run the signed container image, which adds the full gate with ast-grep, gitleaks and zizmor:

docker run --rm -v "$PWD:/workspace" ghcr.io/guardianvigil-labs/prufgate:1.4.0          # the fast gate on this repository
docker run --rm -v "$PWD:/workspace" ghcr.io/guardianvigil-labs/prufgate:1.4.0 prufgate check .

Or add the GitHub Action to a pull-request workflow, pinned to the commit SHA of the v1.4.0 tag (full example):

- uses: actions/checkout@<commit-sha>
  with: { fetch-depth: 0, persist-credentials: false }
- run: python3 -m pip install "PyYAML>=6.0"
- uses: GuardianVigil-Labs/Prufgate@<commit-sha-of-v1.4.0>

Every release is signed and carries build provenance and an SBOM: Verify a release.

Or add it to your pre-commit configuration, .pre-commit-config.yaml (details):

repos:
  - repo: https://github.com/GuardianVigil-Labs/Prufgate
    rev: v1.4.0
    hooks:
      - id: prufgate          # changed files, on every commit
      - id: prufgate-full     # whole repository, before a push (pre-commit install --hook-type pre-push)

Then, in your project:

prufgate check .                          # engines 1-5 on the whole repository
prufgate check --changed --base origin/main   # only what this branch changed
prufgate init --hooks                     # starter config + coding-agent hooks
prufgate baseline                         # record existing findings; only new ones are reported

Exit codes are the same everywhere: 0 clean, 1 blocking findings (P0/P1), 2 PrufGate could not run (bad config, bad base ref, unreadable file, engine crash). A crash never reads as a pass.

Ways to run it

Want

Use

A quick check while coding

prufgate check --changed

What a finding means and how to fix it

prufgate explain <rule> (or --list for every rule): works offline

Adopt it on a codebase that already has findings

prufgate baseline, then commit .prufgate/baseline.yaml: existing findings are recorded and only new ones are reported (details)

Findings inline in VS Code

The VS Code extension (source): runs prufgate check --changed on save, trusted workspaces only (install from a .vsix for now)

One engine

prufgate test-audit, prufgate bola, prufgate review

The full gate with SARIF and Markdown reports

./prufgate.sh fast (changed files) or ./prufgate.sh full, from a checkout. Reports go to reports/prufgate.sarif, reports/prufgate-review.md, reports/gl-code-quality-report.json (GitLab Code Quality) and reports/prufgate-junit.xml.

A pull-request gate on GitHub

The composite action in this repository: docs/CI_INTEGRATION.md

The result where reviewers look

The action writes a short summary (verdict, top findings, what the change touched) to the job summary, and with pr_comment: 'true' keeps it as one pull request comment: details

A coding agent sent back to fix its findings before it stops

prufgate init --hooks (Claude Code Stop hook, which asks once; Cursor rule; the pre-commit hook and the CI gate enforce)

A coding agent stopped from skipping hooks, force-pushing or deleting tests and CI config

prufgate init --hooks (Claude Code, Cursor) or --agent copilot|codex|gemini: fail-closed hooks running prufgate hook <agent> <event> (support table)

Base-branch tests run against the head code (executes repository tests: sandbox or CI only, opt-in)

prufgate verify --differential --allow-run-tests

Mutation smoke test on the changed lines with an installed mutmut, Stryker, PIT, gremlins, go-mutesting or cargo-mutants (executes repository tests: opt-in)

prufgate verify --mutation --allow-run-tests

PrufGate as a Claude Code plugin (hooks plus a /prufgate:prufgate-check skill)

pipx install prufgate, then in Claude Code /plugin marketplace add GuardianVigil-Labs/Prufgate and /plugin install prufgate@prufgate (details)

Other agents and AI reviewers (Cursor Bugbot, Copilot review) told to use PrufGate's findings

prufgate init --agents-md (AGENTS.md) and --bugbot (.cursor/BUGBOT.md): a marked section, updated in place on rerun (details)

An AI second opinion on PrufGate's own low-confidence findings only (opt-in, sends a few lines of code to a provider you choose)

ai_adjudicate: true plus explicit --ai-adjudicate consent: see AI adjudication

Tools for an agent over MCP

prufgate mcp: ten read-only tools, including prufgate_explain

A service that scans on webhook

prufgate api: bearer token, loopback by default

The full command and option reference is in docs/GETTING_STARTED.md, and every configuration key is in docs/CONFIGURATION.md.

Configuration in one look

PrufGate works without configuration. To tune it, add .prufgate/config.yaml:

version: 1
packs: [core, web-ui, ai-safety]      # cti is an opt-in domain pack
exclude: ["dist/**", "vendor/**"]
auth_guards: [requireUser, get_current_user]
tenant_keys: [organization_id, account_id]
exempt:
  src/legacy/adapter.ts:
    route-canned-fallback: "#104: legacy fallback until the payments API migration"
known_debt:
  stub-click-handler: 0                # a ceiling that can only go down

In a pull request (any run with a base ref) PrufGate reads this file as it is on the base branch, so configuration changes take effect after merge: a change cannot exempt its own findings.

AI adjudication

PrufGate is deterministic by default and sends nothing anywhere. Two opt-in modes use an AI provider you choose (a claude CLI, OPENAI_API_KEY, ANTHROPIC_API_KEY or OLLAMA_HOST; --ai-provider names another):

  • prufgate review --ai asks the model to review the whole diff.

  • ai_adjudicate: true in the trusted configuration plus explicit check --ai-adjudicate consent asks one yes/no question per low-confidence finding (medium too if configured): is this real? A local gate can use PRUFGATE_ALLOW_EXTERNAL_AI=1; the Action uses allow_external_ai: 'true'. Configuration alone sends no code. The prompt holds the rule, the finding message and a bounded code window; windows with detected credentials are skipped.

An adjudication verdict can only hide noise. A finding judged not real stops blocking and moves to an "AI-dismissed" list with the model's reason: a section of prufgate check output and of the Markdown report, and a SARIF suppressions entry of kind external (AI-dismissed: <reason>). High-confidence findings are never sent; nothing is ever added or raised. A provider error, a timeout, a reply that is not exactly {"real": true|false, "reason": "..."}, a run over its budget, or code that addresses the judge (a comment naming the AI, LLM, judge, reviewer or PrufGate) keeps the finding as it would be without AI. Verdicts are cached outside the repository. Settings: docs/CONFIGURATION.md.

Documentation

Contributing and security

Contributions are welcome: see CONTRIBUTING.md and the maintainer guide AGENTS.md. Please report vulnerabilities privately, as described in SECURITY.md.

PrufGate is published by GuardianVigil under the Apache License 2.0. The name and logo are trademarks: see TRADEMARKS.md.

Available Tools

10 tools
prufgate_anti_fabricationC
Read-onlyIdempotent

Engine 1: canned fallbacks, simulated return values, stub handlers and other fabricated production logic.

Inputs: base_ref, changed_only, workspace_path.

prufgate_bola_checkB
Read-onlyIdempotent

Engine 5: routes and queries without auth or tenant scoping; the live two-token probe runs only when target_url is set.

Inputs: target_url, workspace_path.

prufgate_contract_checkC
Read-onlyIdempotent

Engine 4: undeclared or hallucinated imports, undocumented environment variables and frontend/backend route mismatches.

Inputs: base_ref, changed_only, workspace_path.

prufgate_diff_guardC
Read-onlyIdempotent

Engine 3: skipped tests, suppression comments, unsafe casts and weakened lint/CI config added in the diff.

Inputs: base_ref, workspace_path.

prufgate_explainA
Read-onlyIdempotent

What a rule catches, how to fix it, and how to record a false positive. Pass the rule ID a finding names.

Inputs: rule (required).

prufgate_fast_scanA
Read-onlyIdempotent

Run engines 1-5 (anti-fabrication, test integrity, diff guard, contract guard, static BOLA) on changed files.

Inputs: base_ref, workspace_path.

prufgate_full_scanA
Read-onlyIdempotent

Run engines 1-5 on the whole workspace plus the diff reviewer; adds the live BOLA probe only when target_url is set.

Inputs: base_ref, target_url, workspace_path.

prufgate_reviewC
Read-onlyIdempotent

Engine 6: deterministic review of the diff against base_ref (default: resolved like prufgate check --changed).

Inputs: base_ref, workspace_path.

prufgate_summaryA
Read-onlyIdempotent

Read the last SARIF/Markdown report written by ./prufgate.sh without running a scan.

Inputs: workspace_path.

prufgate_test_auditC
Read-onlyIdempotent

Engine 2: zero-assertion tests, tautologies, self-mocked units and weakened assertions.

Inputs: base_ref, changed_only, workspace_path.

Tool Schema Changelog

Recent tool additions, removals, and schema changes observed during successful MCP inspections.

Loading tool schema changelog…

    TDQS

    B3.4/5.0

    Scored across 10 tools

    Disambiguation4/5

    Each tool has a distinct role: individual engines, composite scans, rule explanation, and report reading. There is minor overlap between fast_scan/full_scan and the engine-specific tools, but the descriptions clarify scope as changed files vs whole workspace vs single engine.

    Naming Consistency5/5

    All ten tools use the same prufgate_ prefix and snake_case formatting, with no mixed conventions. Whether the name is action-oriented or an engine label, it follows the same predictable namespace pattern.

    Tool Count5/5

    Ten tools map cleanly to six engines plus aggregate scans, explanation, and summary reporting. The count is well-scoped, with no obvious filler or redundant tools.

    Completeness4/5

    The surface covers individual engines 1-6, aggregate scans, rule explanation, and summary report reading. Minor gaps remain around programmatic findings access, false-positive recording, and rule listing, but core scanning workflows are covered.

    Maintenance

    ActivityNo data
    ResponsivenessOutdated data

    Related MCP Connectors

    Related MCP Servers

    • A
      license
      A
      quality
      D
      maintenance
      Read-only MCP tools for coding agents to audit deployment targets, detect fabricated code, review backlog, database schema, analytics, ML models, architecture docs, and decision lenses.
      11
      13 npm
      MIT
    • A
      license
      Not graded
      quality
      C
      maintenance
      Exposes repository reconnaissance tools — file listing and verbatim reading, regex search, dependency-tree parsing across Python/Node/Rust/Go, and test-coverage ingestion — so agents can gather line-numbered evidence from real source files. This lets an audit pipeline physically re-verify every claim and drop findings that cannot be located in the code.
      MIT
    • A
      license
      B
      quality
      C
      maintenance
      Enables AI coding agents to perform evidence-based, read-only audits of supported software projects by exposing bounded project inventory and static-analysis findings for discovery, code review, integration tracing, security, and tests. Reports are returned as Markdown or JSON directly in the tool response, with no code executed or files written.
      8
      17 npm
      MIT