io.github.costajohnt/oss-autopilot

Built and used daily by costajohnt — 3rd biggest contributor to Ink (the React CLI framework behind Claude Code, Gemini CLI, and Codex — 32k+ stars) and repeat contributor to Homebrew.

What It Does

OSS Autopilot monitors all your open PRs across GitHub, alerts you when maintainers leave feedback, helps you draft responses, diagnoses CI failures, and finds new issues matched to your contribution history. It's the workflow engine behind the stats above.

Interactive Dashboard

A Preact SPA that auto-opens when you run /oss — PR management, charts, contribution stats, and status overrides. Also available standalone: npx @oss-autopilot/core dashboard serve.

Engineering Highlights

┌──────────────────────────────────────────────────┐
│  Claude Code Plugin Layer                        │
│  8 slash commands (/oss, /oss-search, …)         │
│  7 specialized agents, contribution skills       │
├──────────────────────────────────────────────────┤
│                                                  │
│  ┌──────────────┐  ┌──────────────────────────┐  │
│  │ MCP Server   │  │ Interactive Dashboard     │  │
│  │ @oss-auto-   │  │ @oss-autopilot/dashboard │  │
│  │ pilot/mcp    │  │ Preact + Vite             │  │
│  │              │  │ PR management, charts,    │  │
│  │ 30 tools     │  │ actions                   │  │
│  │ 6 resources  │  │                           │  │
│  │ 4 prompts    │  │                           │  │
│  └──────┬───────┘  └────────────┬─────────────┘  │
│         │                       │                │
│  ┌──────┴───────────────────────┴─────────────┐  │
│  │ Core Library — @oss-autopilot/core         │  │
│  │ PR monitoring, issue discovery, state mgmt │  │
│  │ GitHub API, CLI, structured JSON output    │  │
│  └────────────────────────────────────────────┘  │
│                                                  │
└──────────────────────────────────────────────────┘

Monorepo with three npm packages — pnpm workspaces with each package independently publishable to npm. Core library, MCP server, and interactive Preact dashboard with shared types.

Three deployment models — Claude Code plugin with 7 specialized agents, MCP server for Cursor/Claude Desktop/Codex/Windsurf, and a standalone CLI with --json structured output. Same core, different interfaces.

Deterministic core, AI orchestration layer — Critical logic (PR status classification, CI failure analysis, state management) lives in tested TypeScript, not in prompts. The CLI returns structured JSON that agents consume. CI failures are categorized into a deterministic taxonomy — actionable vs. fork limitation vs. auth gate vs. infrastructure — rather than asking an LLM each time. 2,600+ tests validate the core independently of any LLM.

Production-grade GitHub API integration — ETag-based HTTP caching, automatic rate limit backoff with retries, bounded concurrency pools, and paginated fetching. Handles the full complexity of fork-based contribution workflows: correct diff ranges, squash commit counting, and --head flag handling for cross-fork PRs. Designed to run daily without hitting API limits.

Human-in-the-loop guardrails — Nothing is posted to GitHub without explicit approval. AI drafts responses but the contributor always reviews before sending. Pre-commit review gates catch issues before they reach maintainers. Factual claims in draft comments are verified against the actual diff before presenting to the user.

Modular extraction — Issue discovery and vetting grew complex enough to extract into its own npm package (oss-scout). Connected via a bridge pattern that maps state between the two systems, following the same approach used by the broader ecosystem of extraction-and-reconnect patterns.

Fresh-fetch architecture — PRs aren't stored locally. Every run fetches live data from GitHub's Search API and enriches each PR with CI status, review decisions, merge conflict detection, maintainer comment classification, and checklist completion. No stale data, no sync bugs.

Security discipline — State files written with 0o600 permissions, data directory created with 0o700. Concurrent state write protection prevents corruption from parallel runs. Runtime schema validation via Zod on every state file read. XSS prevention tested. Input validation hardened across CLI arguments and API responses.

Automated release pipeline — Conventional commits feed into release-please for automatic versioning and changelogs, with CI/CD publishing to npm on merge. 200+ changelog versions across both packages (core v0.1.0 through v3.x, mcp through v5.x) since the first release in January 2025.

Every feature in the list above was driven by real usage — capacity warnings came from overcommitting, "skip comment when code speaks for itself" came from over-commenting, diminishing returns detection came from spending too long searching. The tool is shaped by the contributions it manages.

Install & Usage

Claude Code Plugin (recommended):

/plugin marketplace add costajohnt/oss-autopilot
/plugin install oss-autopilot@oss-autopilot

Restart Claude Code, then run /setup-oss. Done.

First initialize your GitHub username (one-time setup):

npx @oss-autopilot/core@latest init <your-github-username>

Then add to your MCP client config:

{
  "mcpServers": {
    "oss-autopilot": {
      "command": "npx",
      "args": ["@oss-autopilot/mcp@latest"]
    }
  }
}

The MCP server exposes 30 tools, 6 resources, and 4 prompts — the full OSS Autopilot feature set.

# Run any command directly (uses gh auth token automatically)
npx @oss-autopilot/core daily --json
npx @oss-autopilot/core dashboard serve

# Or install globally
npm install -g @oss-autopilot/core

# Or import programmatically
npm install @oss-autopilot/core

import { runDaily, runSearch, runStatus } from '@oss-autopilot/core/commands';

const digest = await runDaily();
const issues = await runSearch({ maxResults: 10 });

All commands return { success, data, error, timestamp } with --json.

Daily workflow (5 min):

1. Run /oss to see what needs attention
2. Work through critical issues (CI failures, maintainer comments, conflicts)
3. Done for now

Commands: /oss (daily check), /oss-search (find issues), /oss-dashboard (interactive dashboard), /oss-guidelines (per-repo guidelines), /pr-ready (pre-push review loop), /plan-ready (plan review loop), /setup-oss (configure), /oss-help (reference)

By the Numbers

Everything Else

Claude automatically dispatches these based on context:

Agents are available in the Claude Code plugin. MCP and CLI users access the same capabilities through tools and commands.

oss-autopilot stats              # Terminal output
oss-autopilot stats --json       # Structured JSON
oss-autopilot stats --markdown   # Shareable markdown report
oss-autopilot stats --badge      # Shields.io endpoint JSON

Add a live badge to your GitHub profile README:

![OSS Contributions](https://img.shields.io/endpoint?url=https://oss-widgets.vercel.app/api/badge/YOUR_USERNAME)

Embed rich SVG widgets powered by oss-widgets — stats card, recent contributions, and 26-week activity graph. All update hourly and support ?theme=dark.

Configuration is stored in ~/.oss-autopilot/state.json (inside the config field). Run /setup-oss to configure interactively, or use setup --set key=value from the CLI:

Does Claude post comments or push code automatically? No. Claude drafts responses and suggests actions. Nothing is posted to GitHub without your explicit approval.

Where is my data stored? All data lives in ~/.oss-autopilot/ — configuration, PR tracking state, event history, and HTTP cache. The dashboard runs locally at http://localhost:3000. Nothing is sent to external servers beyond GitHub API calls.

Can I use this without Claude Code? Yes. The MCP server (npx @oss-autopilot/mcp) works with Cursor, Claude Desktop, Codex, Windsurf, and any MCP client. The CLI (npx @oss-autopilot/core daily --json) runs standalone. The Claude Code plugin provides the best experience with specialized agents and skills, but all core functionality is available through any path.

How do I update? Plugin: /plugin update oss-autopilot. MCP server / CLI: uses npx @latest by default, so you always get the latest. Your configuration is preserved across updates. See the Changelog for what's new.

Any tips for getting started? Set maxActivePRs to 3-5 when starting out. Fewer active PRs with fast responses beats many stale ones. Run /oss every few days — stale PRs are hard to revive.

GitHub CLI authentication errors:

brew install gh    # macOS
gh auth login

Build fails on first run:

# Find your plugin directory
find ~/.claude/plugins -name "oss-autopilot" -type d

# Rebuild
cd <path-from-find-command>/packages/core
npm install
npm run bundle

PRs not showing up:

• Run /setup-oss to ensure your GitHub username is configured
• Check that gh auth status shows you're authenticated
• The plugin only tracks PRs you authored

git clone https://github.com/costajohnt/oss-autopilot.git
cd oss-autopilot
pnpm install                 # Install all workspace dependencies
pnpm test                    # Run all tests across all packages
pnpm start -- daily --json   # Run CLI via tsx (no bundle needed)
pnpm run bundle              # Rebuild CLI bundle (esbuild)

Project structure:

├── commands/                    # 8 plugin slash commands (/oss, /oss-search, /pr-ready, …)
├── agents/                      # 7 specialized agents (PR responder, issue scout, etc.)
├── skills/                      # Contribution best practices
├── workflows/                   # Delegated logic loaded by commands on demand
├── hooks/                       # Plugin hooks (session-start + PreToolUse guards)
├── packages/
│   ├── core/                    # @oss-autopilot/core — CLI + core library
│   │   ├── src/commands/        # CLI subcommands
│   │   ├── src/core/            # Domain logic + tests
│   │   └── dist/cli.bundle.cjs  # Built bundle (auto-generated)
│   ├── mcp-server/              # @oss-autopilot/mcp — MCP server
│   │   └── src/                 # Tools, resources, prompts, server
│   └── dashboard/               # @oss-autopilot/dashboard — Interactive UI
└── pnpm-workspace.yaml          # Workspace definition

Test as a local plugin:

claude --plugin-dir ./oss-autopilot

The plugin includes a built-in pre-commit-reviewer agent that reviews all code changes before committing. For enhanced parallel review, install the pr-review-toolkit plugin (search for it in the Claude Code plugin marketplace) — it adds 5 reviewers that run simultaneously, plus a conditional type-design-analyzer for TypeScript diffs:

Without pr-review-toolkit, the built-in pre-commit-reviewer handles all review phases as a single agent with the same fix-and-re-review loop.

Limitations

• GitHub only — GitLab, Bitbucket, and other forges are not supported. Contributions welcome.
• 1,000 PR cap — GitHub's Search API returns at most 1,000 results per query. If you have more than 1,000 open, merged, or closed PRs, the oldest results may be truncated.
• Individual contributor focus — Designed for solo contributors managing their own PRs. No team dashboards, shared state, or multi-user workflows.

How It Decides

Two docs explain the heuristics that shape which repos surface in discovery and how they're evaluated:

• Repo scores — two distinct 1–10 numbers per repo: the cached history score (your own merged/closed PR outcomes, recency, responsiveness, hostility signals; the default minRepoScoreThreshold of 4 excludes repos below the cutoff from search results) and the fresh health score (repo-vet's weighted rubric over the repo's current activity, PR speed, merge rate, guidelines, and stability).
• Anti-LLM policy detection — scans CONTRIBUTING / CODE_OF_CONDUCT / README for language indicating the project doesn't accept AI-assisted contributions. Hard skip when matched.

Both docs explain the exact rules so you can understand why a given repo did or didn't surface.

API Documentation

Full API documentation for @oss-autopilot/core is available at jcosta.tech/oss-autopilot.

Contributing

Bug fixes, new agents, CLI improvements, and documentation are all welcome. See CONTRIBUTING.md for setup instructions.

License

MIT