# RepoPilot

> Evidence before a coding agent acts, and proof receipts after it changes a project. RepoPilot evaluates public repositories, exact npm versions, bounded project compatibility, and caller-asserted local checks. It is decision support, not a certification, install approval, or merge approval.

## Hosted MCP server

- Endpoint: `https://repopilot.app/api/mcp`
- Transport: Streamable HTTP
- Public use requires no API key.
- All tools are read-only and nondestructive. Six evidence lookups are idempotent.
  `plan_repo_task` creates a fresh proof contract, and contract-backed
  `check_change_risk` records an attempt; do not automatically retry either call.
- Setup and examples: https://repopilot.app/mcp

The eight hosted tools are:

1. `check_dependency` — lean repository-level adoption evidence for one public GitHub repository or npm package.
2. `evaluate_dependency_change` — evaluate an exact npm add/upgrade against a bounded project snapshot before installation.
3. `verify_dependency_change` — verify the resolved version, lockfile receipt, advisory delta, integrity, and caller-asserted command receipts afterward.
4. `analyze_repo` — concise repository brief for an unfamiliar public repository.
5. `plan_repo_task` — task- and full-SHA-aware file orientation plus a short-lived proof contract.
6. `get_artifact` — opt-in full cached repository orientation artifact.
7. `compare_repos` — compare exactly two repositories or npm packages using shared evidence semantics.
8. `check_change_risk` — prioritize review for a pull request, raw diff, or source-free changed-file facts; task proofs may be attached.

Every MCP result returns canonical `structuredContent` plus JSON text for older clients. Preserve `agent.confidence`, `agent.evidence_gaps`, `agent.next_actions`, and evidence provenance. `live_fallback`, `not_analyzed`, or confidence `none` means limited or not verified, never safe.

## Exact npm proof loop

Use `evaluate_dependency_change` before editing a manifest or lockfile. Supply the exact target version plus bounded facts such as package manager, Node version, direct/relevant-peer versions, script names, relative manifest/lock paths, and a lockfile SHA-256 receipt. Preserve the returned `verification_context` unchanged.

After the approved change and local checks, call `verify_dependency_change` with the original context, resolved dependency changes, the new lockfile receipt, and exact command/exit-code receipts. RepoPilot checks those receipts for consistency but did not execute them; `receipt_trust` is `caller_asserted`.

Repository evidence, version-specific npm/deps.dev evidence, project-snapshot claims, and caller-asserted local receipts are separate trust classes. A favourable recommendation means no blocking signal was found in supplied evidence. It does not prove that a package is safe, compatible with every runtime, free of malicious code, or ready for production.

## Task proof loop

Before editing an unfamiliar repository, call `plan_repo_task` with the public repository, task, and full checked-out Git SHA. It returns ranked files, consumers, proof obligations, SHA relation, and an expiring contract using bounded deterministic matching without retaining raw task text.

After editing, call `check_change_risk` with the contract, base/head SHAs, proof statuses, and source-free file facts when possible. The risk score is a deterministic change-shape heuristic, not a defect predictor or merge approval.

## REST APIs

- API guide: https://repopilot.app/api-docs
- `GET /api/v2/check?package=npm:express` — repository-level dependency evidence.
- `GET /api/v2/check?repo=owner/repo` — repository-level evidence for non-npm ecosystems.
- `POST /api/v1/dependency-change/evaluate` — exact npm evaluation.
- `POST /api/v1/dependency-change/verify` — resolved-change and receipt verification.
- `GET|POST /api/v1/onboard?repo_url=https://github.com/owner/repo` — full cached repository brief and artifacts.
- `GET|POST /api/v1/change-risk` — pull-request, raw-diff, or structured-file change risk.

Existing `/api/v1/check` callers retain legacy `go` / `caution` / `avoid` values. The v2 repository-level vocabulary is `proceed_with_review` / `review` / `avoid`.

## Verdict pages and badges

Public cached repository verdict pages use `/r/{owner}/{repo}`. Verdicts are deterministic summaries of available maintenance, ownership, license, CI/test, and known-vulnerability evidence at analysis time.

- **GO** — repository-level evidence supports proceeding to exact-version, compatibility, provenance, and local verification. It does not mean “safe to depend on.”
- **WAIT** — concerns or incomplete evidence require review.
- **AVOID** — material repository-level concerns require explicit approval and alternatives should be considered.

Badge form:

```markdown
[![RepoPilot verdict](https://repopilot.app/api/badge/{owner}/{repo})](https://repopilot.app/r/{owner}/{repo})
```

## Privacy and retention

- RepoPilot analyzes public repositories; GitHub sign-in requests profile identity, not private-repository access.
- Exact project snapshots, raw task text, raw diffs, command strings, proof output, manifest bodies, and lockfile bodies are not persisted.
- The privacy-first local companion uses bounded npm/Git facts and sends relative paths/status/line counts instead of source or diff hunks by default.
- Filenames and public `owner/repo` slugs are still metadata. Inspect the companion's `--dry-run` output and omit repository identity when needed.

Privacy policy: https://repopilot.app/privacy

## Limits and failure semantics

- Per-minute abuse controls may return `429` with `Retry-After`; do not loop.
- Cold live repository fallback is globally capped; cached analyses and short fallback-cache hits continue to serve.
- `503` means fresh analysis capacity is unavailable, not that a target passed.
- Full briefs/artifacts may return `not_analyzed` until a human generates the cached analysis.
- Unknown evidence stays unknown. Absence is claimed only when the relevant evidence source was complete.

## Contact

hello@repopilot.app