Turn your AI coding assistant into a multi-agent team.
Website · Docs · Quick Start · Architecture · Contributing
Works with GitHub Copilot, Cursor, Claude Code, and OpenCode.
One command sets up specialized agents that decompose tasks, work in parallel, and verify each other's output.
One command. Any repo. Any IDE.
npx opencastle init
The CLI asks about your IDEs and stack. It installs agents, skills, and MCP servers tailored to your project.
You can select multiple IDEs and tools — the output is adapted for each one.
| IDE | Output |
|---|---|
| VS Code | .github/ — agents, skills, workflows, prompts |
| Cursor | .cursorrules + .cursor/rules/ |
| Claude Code | CLAUDE.md + .claude/ |
| OpenCode | AGENTS.md + .opencode/ + opencode.json |
MCP servers are auto-configured for your stack in each IDE's native format.
| Command | Description |
|---|---|
opencastle init |
Set up agents in your project |
opencastle update |
Update framework files (keeps your customizations) |
opencastle eject |
Remove the dependency, keep all files |
opencastle destroy |
Remove ALL OpenCastle files (reverse of init) |
opencastle plan |
Generate a convoy spec from a task description |
opencastle run |
Run the Convoy Engine (deterministic, crash-recoverable orchestrator) |
opencastle dashboard |
Open the observability dashboard |
opencastle doctor |
Validate your setup and surface issues |
Add --dry-run to any command to preview what it would change without writing files.
Specialist Agents. Developer, UI/UX, Database, Security, Testing, Reviewer, and more.
On-Demand Skills. Loaded on demand to keep context windows lean. Auto-selected during init based on your stack.
Workflow Templates. Features, bug fixes, data pipelines, security audits — reproducible execution templates.
Quality gates. Fast review after every step. Panel majority vote for high-stakes changes. Lint, test, build checks.
Cost-aware routing. Picks the right model tier (Premium → Quality → Standard → Fast → Economy) based on task complexity.
Self-improving. Agents capture lessons and graduate them into permanent instructions.
npx opencastle dashboard
Visualizes real agent data from your project — sessions, success rates, model usage, execution logs, and panel reviews.
Reads from the same NDJSON logs your agents already write. No setup needed.
📖 Dashboard CLI documentation →
npx opencastle doctorRuns multiple health checks — manifest, configs, skills, observability logs, IDE settings, MCP setup, and environment variables. Useful in CI or after upgrading.
A deterministic, crash-recoverable orchestrator inspired by Steve Yegge's Gas Town. Define tasks in YAML, run overnight, resume after crashes.
npx opencastle run convoy.yml# convoy.yml
name: "Overnight feature batch"
version: 1
adapter: claude-code
branch: feat/reviews
tasks:
- id: migrate-db
agent: database-engineer
prompt: "Create a reviews table migration."
- id: build-component
agent: ui-ux-expert
prompt: "Build a ReviewCard component."
- id: wire-page
agent: developer
prompt: "Add reviews to the place detail page."
depends_on: [migrate-db, build-component]
gates:
- npm run lint
- npm run test- Crash-safe — SQLite WAL persistence survives crashes, power loss, OOM kills. Resume with
--resume. - Isolated — each worker runs in its own git worktree. Changes merge back in dependency order.
- Observable — real-time dashboard auto-starts during execution.
- Multi-runtime — mix Copilot, Claude Code, Cursor, and OpenCode in the same convoy.
Generate a convoy spec from a plain text description — no YAML by hand:
npx opencastle plan --file task.txt📖 Full Convoy Engine documentation →
See ARCHITECTURE.md for the full diagram, workflow templates, and quality gates.
- Fork the repo
- Create a branch —
feat/your-featureorfix/your-fix - Make changes and ensure
npm run build:clipasses - Open a PR
For large changes, open an issue first.
OpenCastle is free and open-source.
For corporate sponsorship inquiries, open a GitHub Discussion.
MIT — Filip Mares, 2026
