PR Context Engine

PR Context Engine is an MCP server that gives AI PR reviewers repository memory. It indexes pull requests, file diffs, review comments, and repository guidelines, then retrieves relevant context during review.

What It Does

Most AI PR reviews only see the current diff. This server adds repository memory:

• Finds similar past PRs and review decisions.

• Retrieves repository rules from docs like README.md and CONTRIBUTING.md.

• Returns structured review context through MCP tools.

• Runs locally by default, with optional Pinecone for team/shared memory.

Quick Start

npm install
npm run build
npm run inspector

By default, the server uses a local JSON vector store at .pr-context-engine/vector-store.json. Users do not need a Pinecone key for public repositories. Add GITHUB_AUTH_TOKEN only if you need private repositories or higher GitHub API limits.

You can also run the MCP server directly after build:

node dist/index.js

Optional Pinecone Mode

Use Pinecone when you want a shared/team vector store instead of local disk.

VECTOR_STORE=pinecone
PINECONE_API_KEY=your_key
PINECONE_INDEX_NAME=pr-context-engine

The Pinecone index must use dimension 384 because the default embedding model is Xenova/all-MiniLM-L6-v2.

Recommended User Flow

1. Run get_server_status to confirm the active vector store.
2. Run index_repo_guidelines for the target repository.
3. Run index_pr on 10-50 important merged PRs from the project history.
4. Run analyze_pr on a new PR.
5. Use search_similar_prs for focused questions like "auth middleware change" or "database migration".
6. Use delete_pr_index before reindexing a stale or changed PR.

MCP Tools

• index_pr: indexes PR metadata, changed-file patches, and review comments.
• index_repo_guidelines: indexes repository docs such as README.md, CONTRIBUTING.md, and custom paths.
• search_similar_prs: semantically searches indexed PR memory.
• analyze_pr: returns current PR summary, diff snippets, relevant guidelines, similar PRs, and recommendations.
• delete_pr_index: removes indexed vectors for a PR before reindexing or cleanup.
• get_server_status: shows active vector store, embedding model, and available tools without exposing secrets.

Recommended First Test

1. Run npm run build.
2. Run npm run inspector.
3. Call get_server_status to confirm local or Pinecone mode.
4. Call index_repo_guidelines with a public repo.
5. Call index_pr for one merged PR from that repo.
6. Call analyze_pr on another PR and inspect returned context.

Retrieval Evaluation

Create an eval file using examples/retrieval-eval.example.json, then run:

npm run eval:retrieval -- examples/retrieval-eval.example.json 5

The command reports recall_at_k, hits, expected PRs, and top retrieved results. This is how you measure whether retrieval quality is improving.

Claude Desktop Example

After npm run build, add a server entry like this to your MCP client config:

{
  "mcpServers": {
    "pr-context-engine": {
      "command": "node",
      "args": ["C:/path/to/PR_Context_Engine/dist/index.js"]
    }
  }
}

For public repositories, this works without secrets. For private repositories, set GITHUB_AUTH_TOKEN in your environment before launching the MCP client.

Production Notes

The local store is best for single-user local MCP usage. For teams, use Pinecone or another shared vector backend, monitor API failures, rotate tokens, and keep the local .pr-context-engine directory out of source control.

See docs/MCP_DEPLOYMENT_GUIDE.md for step-by-step MCP setup and docs/PRODUCTION_GUIDE.md for release, security, evaluation, and operations guidance.