Crew is an Oh My Pi (OMP) extension backed by a durable, repository-scoped local daemon. OMP stays the brain — task intake, scheduling, worker selection, approvals, merge decisions, synthesis. Crew is the hands: it supervises worker processes, speaks harness adapter protocols, persists a durable event journal, recovers after crashes, and feeds display backends.

Everything is delivered through the OMP marketplace (extension + skills, git-cloned from this repository) plus a crewd daemon binary downloaded on demand as a verified GitHub Release asset — no OMP fork, no private APIs, no npm publication.

Why Crew?

Multiagent automation is hard. Most frameworks either:

• Put all the intelligence in the agent (risky, hard to debug, no recovery)
• Put all the intelligence in your code (complex, brittle, no replay)

Crew splits the difference: OMP decides what to do, Crew ensures it happens and can be replayed.

Key benefits:

• Durable event journal. Every action is persisted before it executes. Crash? Replay from the last known state.
• Redaction by construction. Secrets never reach the journal. Enforced at the type level.
• SQLite-backed. Query your automation history with SQL. No proprietary format.
• Adapter-agnostic. Claude, Codex, Copilot, OMP-RPC — plug in any worker.
• No model calls required for monitoring. Check runtime status, task state, or run history without spending a token.

If you're building multiagent systems that need to be auditable, recoverable, and debuggable, Crew is your foundation.

Installation

Crew consists of two components, installed in two steps:

• Plugin: the OMP extension + skills, pulled from this repository via the OMP marketplace
• Binary: the crewd runtime daemon, downloaded as a verified GitHub Release asset

/marketplace add nikolasd/crew
/marketplace install crew@crew

Exit and start a new omp session. /reload-plugins does not reload extension modules, so /crew-install (and every crew_* tool) only exists once a fresh session has loaded the installed module. Then:

/crew-install
/crew health

This repository is private. The marketplace step git-clones it, so you need your own GitHub read access to nikolasd/crew — an SSH key registered with GitHub, or a gh auth login session backed by a git credential helper. /crew-install additionally needs a GITHUB_TOKEN or GH_TOKEN environment variable set, or that same gh auth login session, to download and verify the release asset. The binary is cached under your Crew state root.

Once installed, docs/plugin-usage.md is the user manual: every tool and command the extension registers, and the recommended flow for running a task through it.

To uninstall:

/marketplace uninstall crew@crew

Development

For contributors building or modifying Crew itself (not for end users — see Installation above). docs/getting-started.md is the developer manual — start there for the full setup/build/test/config walkthrough.

Prerequisites: Bun 1.3.14+, macOS or glibc Linux on arm64/x64, and Rust — via rustup (recommended: automatically respects the pinned 1.97.1 in rust-toolchain.toml) or your system package manager. For the full OMP integration you also need OMP ≥ 17.0.7.

git clone https://github.com/nikolasd/crew.git
cd crew
bun run setup               # installs JS deps + builds the crewd runtime
bun run check               # schema drift check + build + all tests

To exercise the extension against your local changes before opening a PR, load it from its source path directly:

OMP_CREW_BINARY="$PWD/target/debug/crewd" \
  omp --extension ./packages/extension/src/index.ts

Ask the model to use crew_task, crew_worker, and crew_run, then open /crew to watch runs live. See docs/plugin-usage.md for the full tool reference and docs/manual-testing.md for the full walkthrough. For running crewd directly instead of through OMP, see docs/cli-reference.md.

packages/extension/dist/index.js is committed to git and verified in CI (a bundle-check job rebuilds and diffs it), since it's the entry point the marketplace-installed plugin loads. Any change under packages/extension/src/ must be followed by bun run build and committing the rebuilt bundle.

Platform note: the bundle embeds Bun's platform-specific module shim, so a rebuild on a different platform (e.g. macOS/arm64) does not byte-match CI's bundle-check — observed with Bun 1.3.14, where a darwin-arm64 rebuild diverges from CI's linux-x64 output. Always refresh the committed bundle via the refresh-bundle workflow (builds on linux-x64 + pinned Bun and uploads the artifact to commit), or a linux-x64 build, or CI will reject the commit as a stale bundle.

Contributing

Contributions are welcome. Before submitting a PR:

1. Read docs/getting-started.md and docs/architecture.md.
2. Run bun run check — schema drift, build, and all tests must pass.
3. Follow the Non-Negotiable Invariants — changes that weaken them will be rejected.
4. Use descriptive commit messages. Reference issue numbers when applicable.
5. Describe what changed and why, link related issues, and request review. (There is no PR template to fill out — just write a clear description.)

For detailed guidelines, see CONTRIBUTING.md. For the release/publishing process, see CONTRIBUTING.md's Releasing section.

Author

Crew was created by Nikolas Demiridis as part of the Oh My Pi ecosystem.

For questions, issues, or contributions, please open a GitHub Issue on this repository.

License

This project is licensed under the MIT License. See the LICENSE file for full terms.

Known Limitations

This is a pre-1.0 project. The review backlog is empty — the one open item is an unreproduced test-flake watch, tracked in the maintainer's local, gitignored REVIEW.md (resolution history lives in docs/journal.md). What remains below are environment and protocol walls, verified against the current codebase. Every adapter is installed and authenticated here, and live TUI conformance runs against all four — claude, codex, and omp-rpc are fully green, while copilot alone is blocked on a confirmed vendor monthly-quota wall (raw reports under release/live-conformance/ with an erratum). None of the below is a "requires a vendor CLI" caveat.

• ACP v1 has no durable session handle, so Copilot cannot resume across processes. A session that completed a real turn answers session/load with Resource not found, which fails session_resume and runtime_restart. A protocol wall, not an adapter defect.
• ACP v1 exposes no subagent-observation variant, so Copilot's vendor-side delegation cannot be normalized to NestedWorkerObserved. Pending a newer ACP version.
• Copilot's turn-dependent scenarios are unprovable on this account's monthly quota. The tailed session's own event log records a typed session.error (errorCode: quota_exceeded, statusCode: 402) — independently confirmed, not inferred from a timeout. Refilling the workspace makes the two remaining scenarios provable with no code change.

Consciously deferred features, each with a decision trigger, live in docs/future-features.md.