opencastle

OpenCastle

Turn your AI coding assistant into a multi-agent team.

Website · Docs · Quick Start · Architecture · Contributing

Works with GitHub Copilot, Cursor, Claude Code, OpenCode, Windsurf, Codex CLI, and Antigravity.

One command sets up specialized agents that decompose tasks, work in parallel, and verify each other's output — interactively through IDE chat or automated via the Convoy Engine.

Use it live in your IDE. Run it overnight with the CLI.

Quick Start

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.

What gets generated

IDE	Output
VS Code	.github/ — agents, skills, workflows, prompts
Cursor	.cursorrules + .cursor/rules/
Claude Code	CLAUDE.md + .claude/
OpenCode	AGENTS.md + .opencode/ + opencode.json
Windsurf	.windsurfrules + .windsurf/rules/
Codex CLI	AGENTS.md + .codex/
Antigravity	GEMINI.md + .agents/

MCP servers are auto-configured for your stack in each IDE's native format.

Project CLI

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 skills	Skill refinement and failure tracking
opencastle package	Package orchestrator as a plugin for IDE marketplaces

Convoy CLI

Command	Description
opencastle start	Go from idea to convoy spec in one command (PRD → validate → convoy → validate → fix)
opencastle plan	Run a single prompt template step (generate PRD, convoy spec, or validate)
opencastle validate	Validate a convoy YAML spec file without executing it
opencastle run	Run the Convoy Engine (deterministic, crash-recoverable orchestrator)
opencastle dispute	Manage convoy dispute resolution

Observability CLI

Command	Description
opencastle dashboard	Open the observability dashboard
opencastle doctor	Validate your setup and surface issues
opencastle agents	Manage persistent agent identities
opencastle baselines	Manage visual regression baselines
opencastle log	Append a structured event to the observability log
opencastle lesson	Append a structured lesson to LESSONS-LEARNED.md
opencastle artifacts	Manage filesystem artifact storage (prune old convoy artifacts)
opencastle insights	Analyze convoy execution history and generate recommendations

Add --dry-run to any command to preview what it would change without writing files.

📖 Full CLI documentation →

What's Inside

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. All 51 skills and plugins scored 100% on the tessl Skill Evaluator.

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.

Hybrid by design. Use OpenCastle interactively through IDE chat for hands-on development, or run automated batch jobs overnight via the Convoy Engine.

Dashboard

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 →

Doctor

npx opencastle doctor

Runs multiple health checks — manifest, configs, skills, observability logs, IDE settings, MCP setup, and environment variables. Useful in CI or after upgrading.

📖 Doctor CLI documentation →

## Convoy Engine

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
branch: feat/reviews

tasks:
  - id: migrate-db
    agent: database-engineer
    prompt: "Create a reviews table migration."
    files: ["db/migrations/"]

  - id: build-component
    agent: ui-ux-expert
    prompt: "Build a ReviewCard component."
    files: ["src/components/reviews/"]

  - id: wire-page
    agent: developer
    prompt: "Add reviews to the place detail page."
    depends_on: [migrate-db, build-component]
    files: ["src/pages/places/"]

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, OpenCode, Windsurf, Codex, and Antigravity in the same convoy.

Generate a validated convoy spec from a plain text description — no YAML by hand:

npx opencastle start --text "Add user reviews to the place detail page"

The start command runs 7 steps automatically: generate PRD → validate PRD → auto-fix PRD if needed → assess complexity → generate convoy spec → validate spec → auto-fix spec if needed. For lower-level control, use opencastle plan to run individual steps.

📖 Full Convoy Engine documentation →

Architecture

See ARCHITECTURE.md for the full diagram, workflow templates, and quality gates.

Contributing

• Fork the repo
• Create a branch — feat/your-feature or fix/your-fix
• Make changes and ensure npm run build:cli passes
• Open a PR

For large changes, open an issue first.

Support

OpenCastle is free and open-source.

For corporate sponsorship inquiries, open a GitHub Discussion.

License

MIT