About Oss Autopilot
Oss Autopilot is an MCP server for open-source contributors to track GitHub pull requests, discover issues, and manage contributions. It also provides evaluation tools, state resources, and prompts over stdio or Streamable HTTP.
Key features
- 01Track and manage pull requests
- 02Search and assess contribution issues
- 03Post comments and claim issues
- 04Maintain repository guidelines
- 05Provide state resources and prompts
- 06Support stdio and Streamable HTTP
Use cases
- 01Review and prioritize daily PR status
- 02Find suitable GitHub issues to contribute to
- 03Assess PR compliance and repository health
- 04Derive repository guidance from PR feedback
MCP support
As an MCP server, it exposes PR tracking, issue discovery, and contribution-management tools, along with status and configuration resources and related prompts. It supports stdio and Streamable HTTP.
Configuration
Local server config is the same for every client. Copy it into your client’s MCP config file.
npx -y @oss-autopilot/mcpGITHUB_TOKENSecretGitHub personal access token or token from `gh auth token`Local servers need a runtime on your machine (e.g. Node.js 18+). Config file locations differ by client; see your client’s docs.
Tools
30Saved tool list · updated 3w ago
This tool takes no parameters.
Details
From the project README@oss-autopilot/mcp
MCP server for OSS Autopilot — exposes PR tracking, issue discovery, and contribution management as MCP tools for any MCP-compatible client.
What It Provides
| Feature | Count | Description |
|---|---|---|
| Tools | 30 | daily, status, search, features, vet, vet-list, verify-issue, track, compliance-score, repo-vet, strategy, comments, post, claim, config, init, setup, check-setup, startup, dismiss, undismiss, move, state-show, state-sync, state-unlink, guidelines-list, guidelines-get, guidelines-store, guidelines-reset, guidelines-fetch-corpus |
| Resources | 6 | oss://status, oss://config, oss://prs, oss://prs/shelved, oss://pr/{owner}/{repo}/{number}, oss://repo/{owner}/{repo}/guidelines |
| Prompts | 4 | triage (PR prioritization), respond-to-pr (draft response), find-issues (discover issues), extract-learnings (distill per-repo guidance from past PR feedback) |
Supports stdio (default) and Streamable HTTP transports.
Prerequisites
- Node.js 22+
- GitHub CLI authenticated (
gh auth login)
Quick Start
# 1. Add the server to your MCP client (see config examples below)
# 2. From your MCP client, call the `init` tool once with your GitHub username
# (this writes ~/.oss-autopilot/state.json).
# 3. Use the other tools — e.g. `daily` to check your PRs, `search` to find issues.
Client Configuration
Claude Desktop
Add to your claude_desktop_config.json:
{
"mcpServers": {
"oss-autopilot": {
"command": "npx",
"args": ["@oss-autopilot/mcp@latest"]
}
}
}
Cursor
Add to .cursor/mcp.json in your project or global config:
{
"mcpServers": {
"oss-autopilot": {
"command": "npx",
"args": ["@oss-autopilot/mcp@latest"]
}
}
}
Codex CLI
codex mcp add oss -- npx @oss-autopilot/mcp@latest
Windsurf
Add to your Windsurf MCP config:
{
"mcpServers": {
"oss-autopilot": {
"command": "npx",
"args": ["@oss-autopilot/mcp@latest"]
}
}
}
Streamable HTTP (any client)
Run the server in HTTP mode instead of stdio:
npx @oss-autopilot/mcp@latest --http --port 3001
The server listens at http://127.0.0.1:3001/mcp and accepts POST requests.
Authentication
HTTP mode requires a bearer token on every request (stdio mode does not — it inherits the parent process identity). On first --http startup, the server generates a random 32-byte token and writes it to ~/.oss-autopilot/mcp.token with 0600 permissions. The path and "newly generated" marker are logged to stderr at startup.
Every HTTP client must send:
Authorization: Bearer <contents of ~/.oss-autopilot/mcp.token>
Requests without a valid Authorization header return 401. The Bearer scheme is case-insensitive per RFC 7235; the token value itself is compared byte-for-byte in constant time. Requests with a non-loopback Host header return 403 (DNS-rebinding defense). Requests must declare a numeric Content-Length ≤ 1 MiB — missing, non-numeric, or oversize Content-Length values return 413.
The token persists across restarts. To rotate, delete the file and restart — a new token will be generated. To relocate the token (e.g. for multi-user or CI setups), set OSS_AUTOPILOT_MCP_TOKEN_PATH to the desired absolute path before starting the server.
Tools Reference
| Tool | Description | Read-only |
|---|---|---|
daily | Run daily PR monitoring check with prioritized summary | No |
status | Show current PR tracking status | Yes |
search | Search GitHub for contributable issues | Yes |
features | Find feature-scoped opportunities in repos with 3+ merged PRs (relationship-anchored) | Yes |
vet | Analyze an issue for contribution suitability | Yes |
verify-issue | Deterministically verify issue state and linked-PR claims before vetting | Yes |
vet-list | Re-vet all available issues in the curated issue list | No |
track | Inspect a PR: read-only metadata lookup (nothing is tracked or persisted) | Yes |
compliance-score | Score a PR against opensource.guide best practices (#1245) | Yes |
repo-vet | Compute the repo health rubric (1–10 + verdict) for owner/repo (#1271) | Yes |
strategy | On-demand contribution strategy snapshot via computeStrategy() (#1243) | Yes |
comments | Fetch and display PR comments | Yes |
post | Post a comment on an issue or PR | No |
claim | Claim an issue by posting a comment | No |
config | Get or set configuration values | No |
init | Initialize with a GitHub username | No |
setup | Configure preferences (languages, interests) | No |
check-setup | Check if setup is complete | Yes |
startup | Run startup checks (auth, state, config) | No |
dismiss | Dismiss an issue from notifications | No |
undismiss | Re-enable notifications for a dismissed issue | No |
move | Move a PR between states (attention, waiting, shelved, auto) | No |
state-show | Show current state persistence mode (local or Gist) and sync status | Yes |
state-sync | Force push current state to the backing Gist | No |
state-unlink | Disconnect from Gist persistence and switch to local-only mode | No |
guidelines-list | List repos that have stored guidelines (always empty in local mode) | Yes |
guidelines-get | Read per-repo learning guidelines extracted from past PR feedback | Yes |
guidelines-store | Persist per-repo guidelines (8 KB cap; requires Gist mode) | No |
guidelines-reset | Tombstone the guidelines file for a repo | No |
guidelines-fetch-corpus | Fetch raw PR comment bundles for the host's extract-learnings prompt to consume | No |
Resources Reference
| Resource URI | Description |
|---|---|
oss://status | PR tracking status (cached local state) |
oss://config | Current configuration |
oss://prs | Active open PRs from last daily digest |
oss://prs/shelved | Shelved PRs |
oss://pr/{owner}/{repo}/{number} | Detail for a specific PR |
oss://repo/{owner}/{repo}/guidelines | Per-repo learning guidelines (markdown) |
Prompts Reference
| Prompt | Args | Description |
|---|---|---|
triage | none | Fetches daily digest and builds a prioritized triage list |
respond-to-pr | prUrl | Fetches PR comments and context for drafting a response |
find-issues | maxResults? | Searches for issues ranked by viability score |
extract-learnings | repo, corpus, existingGuidelines? | Distills durable per-repo guidance from raw PR comment bundles (#867) |
Programmatic Usage
The server can also be imported and used as a library:
import { createServer } from '@oss-autopilot/mcp';
const server = createServer();
// Connect to your own transport
More Information
See the main repository README for the full documentation, including the Claude Code plugin, CLI usage, dashboard, and contributing guide.
License
MIT
Comments
New to MCP?
From what the protocol is to installing your first server, these answers cover most first questions.
An open protocol that defines how AI applications connect to external tools and data in a uniform way. Clients talk to the model; MCP servers provide tools, resources and prompt templates. They exchange standard messages, so one server can be reused by many clients.