repopilot
Local RepoPilot MCP server for reviewing Git changes, auditing repositories, generating context briefs, and explaining analysis results without uploading source.
Review Git changes locally for decisions, findings, blast radius, tiered signals, weakened tests/checks, security boundaries, behavior changes, algorithmic changes, and taint-lite paths; supports base/head refs, scopes, filters, baselines, intent metadata, and gated results.
Audit a repository, folder, or file for architecture, coupling, code quality, security, and testing findings with metrics and risk summary.
Generate a budgeted Markdown repository brief for orientation, optionally focused on security, architecture, quality, framework, or all.
Explain how RepoPilot treats a file, rule, or signal; why a scan or review finding fired; and details of a review signal such as confidence, affected files, verification, and limits.
Run locally from MCP-compatible agents; scan, context, and explain operations are read-only, and optional review checks require explicit one-time user consent.
Provides setup steps, an MCP entry, and an AGENTS.md snippet for GitHub Copilot's coding agent and other agents, enabling access to RepoPilot's local scan and review tools.
RepoPilot
Local, deterministic review for Git changes.
RepoPilot helps developers and teams inspect a change before merge. It reports structural evidence about security boundaries, behavior, local imports and exports, and the files affected through the dependency graph. The same review can run from a terminal, in CI, or through an agent integration.
The analysis runs on the machine or CI runner that invokes it. It does not send source to a hosted service or call an embedded language model. Findings point to code and explain what to check; reviewers still decide whether a change is safe for their application.
Example: a change that weakens its own tests
This change drops a range check from applyDiscount. The same change skips the
test that would now fail, removes an assertion from another test, and lets the
CI test step fail without failing the job. Every check that still runs passes.
repopilot review . lists them right after its decision:
Decision: REVIEW (Change Proof: REVIEW)
...
Checks this change weakened (details under Review signals):
⚑ check gate relaxed — .github/workflows/ci.yml:10
⚑ assertions removed — src/pricing.test.ts:5
⚑ test skipped — src/pricing.test.ts:9Further down, each signal explains itself:
⚑ check gate relaxed — .github/workflows/ci.yml:10 check step `npm test` in job `test` now has `continue-on-error: true`
⚑ assertions removed — src/pricing.test.ts:5 "applyDiscount > takes a percentage off the total": assertions 2 → 1
⚑ test skipped — src/pricing.test.ts:9 `it.skip` added on "rejects a discount above 100%"; its result no longer fails the runThese are excerpts of one run (recording). Replay it with
scripts/demo-weakened-tests.sh <empty-dir>, then run repopilot review <empty-dir>.
Related MCP server: codebase-doctor
Install and review
cargo install repopilot
# or
npm install -g repopilot
repopilot review . --base origin/mainFor uncommitted work, run repopilot review .. The first screen gives one
PASS, REVIEW, BLOCK, or NOT ASSESSED decision, its reasons, coverage
limits, and a next action.
A review can surface:
changes to authentication, request trust, deployment, dependencies, or secret configuration;
added or removed behavior such as network calls, subprocesses, filesystem writes, SQL, or error handling;
changed input-to-sink paths, algorithmic structure, or local import/export contracts;
tests the change stopped running or weakened: committed focus markers such as
it.only, newly skipped tests (it.skip,@pytest.mark.skip,t.Skip,#[ignore]), removed or substituted test cases, tests that lost assertions, new lint, type, or coverage suppressions, and relaxed CI or tool gates (continue-on-error,|| true, lowered coverage thresholds, strict mode off), so a green run cannot hide them;direct dependents and the wider impact of changed files.
Signals are advisory evidence. For example, a taint-lite signal shows that a recognized input can reach a recognized sink in the changed source. Confirm the impact in the context of the application and its configured checks.
Example: review an image-processing change
The Wagtail example removes an authorization check and passes request data to a subprocess in a one-file change. RepoPilot reports both the boundary change and the input-to-process flow.
Replay the example on the pinned test repository:
python3 scripts/zoo.py clone --only wagtail
scripts/demo-agent-edit.sh .zoo/wagtail
repopilot review .zoo/wagtailA reported flow is a path to investigate. RepoPilot does not prove that it is exploitable or that the application is safe.
Use in a team
Gate a branch review on high-confidence signals:
repopilot review . --base origin/main --fail-on-review definitelyA review is VERIFIED only when checks configured for the repository are
explicitly selected with --verify and pass on the reviewed revision. Generate
suggestions for a first setup with
repopilot init --suggestions-output repopilot-suggestions.toml; the file is
separate from active configuration. See configuration.
For a broader repository view, run repopilot scan .. Use
repopilot baseline create . to adopt existing findings before gating new work.
Agent and CI integrations
The same review can check work from a human or a coding agent. Record a starting point and review the changes made since it:
repopilot snapshot
# Work happens here.
repopilot review --since-snapshotWhen the working tree is already dirty, the marker also records a baseline commit of those uncommitted files, so the later review covers only what changed after the snapshot. It shows what changed, not who changed it. See common workflows.
In Claude Code, the RepoPilot plugin runs that loop for every session and stops Claude from finishing while a test it skipped, focused, removed, or weakened is unexplained:
/plugin marketplace add MykytaStel/repopilot
/plugin install repopilot@repopilotCodex installs the same plugin (codex plugin marketplace add MykytaStel/repopilot, then codex plugin add repopilot@repopilot). Gemini CLI installs it as an
extension (gemini extensions install https://github.com/MykytaStel/repopilot),
and Cursor runs it through project hooks. For GitHub Copilot's coding
agent and other agents, RepoPilot provides setup steps, an MCP entry, and an
AGENTS.md snippet. See Guard your agent runs.
RepoPilot also provides a local stdio MCP server and a GitHub Action. The MCP
server gives an agent access to the local scan and review tools. The Action runs
RepoPilot on the Actions runner and can publish SARIF or a pull request summary.
See Guard your agent runs, MCP server,
and GitHub integration. MCP Registry
name: mcp-name: io.github.MykytaStel/repopilot.
More capabilities
repopilot ai context .creates a local handoff with repository facts, findings, and a prioritized plan. It makes no model calls.repopilot initcreates configuration or integration files for review.Reports are available as console, Markdown, JSON, HTML, and SARIF.
Documentation
Contributing and development setup: CONTRIBUTING.md.
License
Licensed under either of Apache License, Version 2.0 or MIT license at your option.
Available Tools
6 toolsrepopilot_contextARead-onlyIdempotent
Generate a budgeted Markdown brief of the repository (risks, hotspots, structure) to read before editing, especially in an unfamiliar codebase. Use it for orientation. For the full findings as JSON, use repopilot_scan; to review a change, use repopilot_review_change. Pass analysis_handle to make sure the brief still matches an earlier scan or review. Built locally from a scan; no AI service is called and nothing is uploaded.
Inputs: path, focus, budget, config, +2 more. Returns: decision, evidence, markdown, change_proof
repopilot_explain_fileARead-onlyIdempotent
Explain how RepoPilot treats one file: the role it assigns (for example test, generated, config, or CLI command handler) and the evidence for it, which rules apply, the ordered overrides that change a rule's severity, and whether the default profile would show the result. Use it to understand a file without a report, or to ask why a rule (rule, optionally one detector signal) would or would not fire there. To explain a finding already in a scan or review, use repopilot_explain_finding; for a review signal, use repopilot_explain_review_signal. Reads local files only and returns JSON.
Inputs: path (required), rule, signal. Returns: object
repopilot_explain_findingARead-onlyIdempotent
Explain why a finding was reported. Use it when you have a finding_id from repopilot_scan or repopilot_review_change in this session and need to know why it fired or whether it still reproduces. It replays the stored rule decision against the current workspace, returns the full decision trace, and reports whether the decision matched or drifted. For a review signal (signal_id), use repopilot_explain_review_signal; for a file with no finding, use repopilot_explain_file. Local-only.
Inputs: finding_id (required), source, line_start, evidence_path, +1 more. Returns: object
repopilot_explain_review_signalARead-onlyIdempotent
Explain one signal from a review: where it came from, its confidence tier, whether it can fail a gate, the files it affects, how to verify it, and its limits. Use it after repopilot_review_change, with a signal_id from that report's tiered_signals. For a finding (finding_id in the findings array), use repopilot_explain_finding instead. Reads the stored review only; it runs nothing and changes nothing.
Inputs: signal_id (required), analysis_handle. Returns: object
repopilot_review_changeADestructive
Review a Git change locally: what it touched and which checks it weakened. Use it before finishing work or before a merge. By default it reviews uncommitted work (working tree vs HEAD); pass base (e.g. "origin/main") to review a branch. For a repository health check that is not about one change, use repopilot_scan instead. Returns a JSON report with a decision, findings on changed lines vs the rest, blast radius (files that import the changed files), and deterministic signals grouped by confidence tier (definitely / maybe / noise) in tiered_signals: checks the change weakened (focused or skipped tests, removed tests, tests that lost assertions, new lint/type/coverage suppressions, relaxed CI or tool gates); security-boundary changes (auth, CORS, CI, dependency manifests, committed .env); behavioral changes (network, subprocess, filesystem, env, dependency, migration, or raw SQL added; error handling, an auth check, or a test removed; a changed quiet-fallback candidate in parsed JS/TS or Python; a removed named TypeScript/JavaScript export that a resolved local caller still imports; a Rust public function whose parameter count changes while a proven module-qualified caller keeps the old argument count); algorithmic changes (deeper nesting, a new nested loop, a grown function, new recursion); and taint-lite reachability (HTTP request or process arguments reaching SQL, exec, filesystem-write, or outbound-network sinks in a changed function). Treat quiet-fallback signals as candidates to verify, not proof of harmful behavior, and preserve intentional recovery paths. Signals are evidence, not verdicts; explain one with repopilot_explain_review_signal. RepoPilot uploads nothing. Only a non-empty verify array selects configured local checks. Each selected check runs only after one-time form-mode Elicitation consent from a client that negotiated protocol 2025-11-25 and elicitation.form; otherwise it is skipped. The prompt shows the command, exact arguments, and working directory. The process still has the host user’s permissions and is not sandboxed; its captured output is bounded and redacted.
Inputs: base, head, path, limit, +11 more. Returns: object
repopilot_scanARead-onlyIdempotent
Audit a repository, folder, or file for findings across architecture, coupling, code quality, security, and testing, and return the JSON scan report (findings, metrics, risk summary). Use it for a health check that is not about one change. To review what a change touched or which checks it weakened, use repopilot_review_change; for a short Markdown brief to read before editing, use repopilot_context. Explain a returned finding with repopilot_explain_finding. Runs entirely on disk; nothing is uploaded.
Inputs: base, path, limit, scope, +4 more. Returns: object
Tool Schema Changelog
Recent tool additions, removals, and schema changes observed during successful MCP inspections.
Loading tool schema changelog…
TDQS
Scored across 6 tools
The three top-level entry points are clearly distinct: repopilot_scan (health audit as JSON), repopilot_review_change (diff review), and repopilot_context (Markdown brief), and each description explicitly says when to prefer the others. The only soft spot is the trio of explanation tools, whose names (explain_finding vs explain_review_signal vs explain_file) are similar and whose correct choice depends on knowing whether you hold a finding_id, a signal_id, or a path — the descriptions do call this out, so misselection is unlikely but not impossible.
All six tools share a repopilot_ prefix and snake_case with a clear verb_ or verb_noun shape (review_change, explain_finding, scan). The only deviation is repopilot_context, which uses a bare noun rather than a verb, breaking the otherwise predictable pattern slightly. Still highly readable and consistent overall.
Six tools is well-scoped for a local repo audit/review server: three do the work (scan, review_change, context) and three provide explanation for the objects those produce. No tool looks like filler, and nothing essential appears to be collapsed into an overloaded god-tool.
The read/analyze/explain lifecycle is covered thoroughly: audit, change review, orientation brief, and explanation for findings, signals, and files. Minor gaps remain — there is no tool to inspect or manage configuration profiles/suppressions or to enumerate available rules/detectors in bulk, which an agent doing suppression cleanup would want. For a read-only analysis server, core workflows are not blocked.
Maintenance
Related MCP Connectors
Risk-scan a diff, flag AI-generated-code tells, find secrets. 5 of 7 tools need no account.
Security reviews for coding agents: diffs checked against your org policy and live infrastructure.
Checks AI-written code for leaked keys and made-up packages, on your machine. Never uploads code.
Audit GitHub repos for malicious and supply-chain code before you depend on them.
Related MCP Servers
- AlicenseAqualityBmaintenanceLocal code-graph engine for AI coding workflows: static analysis of Rust, Cangjie, ArkTS, TypeScript/JavaScript, C/C++, Python, and Shell projects into symbols, imports, call graphs, and quality gates, with pre-commit change review and impact analysis. Read-only by design — never executes build scripts or uploads code.61MIT
- AlicenseAqualityBmaintenanceEvidence-backed, read-only codebase diagnostics for coding agents: secrets, dependency, RLS, and source-impact audits with a verify loop and PR review verdicts. Never writes files or uses the network by default.7146 npmMIT
- AlicenseAqualityBmaintenanceEnables AI assistants to turn local git history into answers about recent activity, file hotspots, change coupling, stale or already-merged branches, contributors, ref comparisons, and drafted release notes. It runs locally over stdio with a git subcommand allowlist, so it is read-only, works offline and without shell access, and is safe to auto-approve.13284 npmMIT
- AlicenseNot gradedqualityBmaintenanceEnables AI assistants to audit git diffs, scan for leaked secrets and unsafe APIs, detect breaking changes, triage issues, generate changelogs, and assess repository hygiene through native stdio JSON-RPC tools.7Apache 2.0