AI agents that build
software together.
OpenCastle turns GitHub Copilot, Cursor, Claude Code, OpenCode, Windsurf, Codex, and Antigravity into coordinated multi-agent development teams. One orchestrator. Multiple specialists. Zero chaos.
Everything you need to orchestrate AI agents at scale.
A hybrid framework for interactive IDE chat and automated batch execution — with instructions, skills, agent definitions, workflow templates, and quality gates battle-tested on a production codebase.
Team Lead Orchestrator
A single coordinator that analyzes, decomposes, delegates, and verifies work across specialist agents — never writes code itself.
Specialist Agents
From frontend developer to security expert, each with curated tools, model tiers, and domain-specific skills.
On-Demand Skills
Modular knowledge files loaded per task. Keeps context windows lean while enabling deep domain expertise. Stack-specific skills are auto-selected during init. All 51 skills scored 100 on the tessl Skill Evaluator.
Workflow Templates
Reproducible execution plans for features, bug fixes, data pipelines, schema changes, security audits, and more.
Quality Gates
Validation gates — from secret scanning and blast radius checks to panel majority vote and final smoke tests. Sensitive files always get reviewed, even for 1-line changes.
Self-Improvement
Agents capture lessons learned from retries and failures, then graduate them into permanent instructions.
Specialized agents, intelligent routing.
Precisely developed specialist agents, each with its own model tier, tool access, and file partition. The Team Lead orchestrator routes tasks based on complexity scoring.
One command. Your stack.
Run npx opencastle init and follow the guided onboarding.
Select your IDEs, tech tools, and team tools.
The CLI installs only the plugins you need.
# Install OpenCastle $ npx opencastle init # Select your IDEs: VS Code / Cursor / Claude Code / OpenCode / Windsurf / Codex / Antigravity # Select Tech Tools: Sanity / Supabase / Prisma / Vercel / Netlify / NX / Turborepo / ... # Select Team Tools: Linear / Jira / Trello / Slack / Teams / Notion # Key commands $ npx opencastle update # Update framework, preserve customizations $ npx opencastle start # Go from idea to convoy spec (PRD → validate → convoy) $ npx opencastle dashboard # View observability dashboard $ npx opencastle doctor # Check system health and configuration # See all commands: docs/cli
Convoy Engine. Crash-safe. Isolated. Observable.
Inspired by Steve Yegge's Gas Town, the Convoy Engine is a deterministic,
crash-recoverable orchestrator backed by SQLite WAL-mode persistence.
Each worker gets its own git worktree — no shared mutable state.
Define your batch in convoy.yml, run it overnight, and resume exactly where you left off.
name: "Overnight feature batch" version: 1 concurrency: 2 adapter: claude branch: feat/reviews defaults: timeout: 30m max_retries: 1 tasks: - id: migrate-db agent: database-engineer prompt: | Create a new migration for a reviews table. timeout: 10m files: ["db/migrations/"] - id: build-component agent: ui-ux-expert prompt: | Build a ReviewCard component. timeout: 15m files: ["src/components/reviews/"] - id: wire-page agent: developer depends_on: [migrate-db, build-component] prompt: | Add reviews to the place detail page. files: ["src/pages/places/"] gates: - npm run lint - npm run type-check - npm run test
Crash-Safe Execution
SQLite WAL-mode persistence means convoy state survives crashes,
power loss, and OOM kills. Resume exactly where you left off
with --resume.
Worktree Isolation
Each worker gets its own git worktree — no shared mutable state. Changes merge back in dependency order once the convoy completes.
Health Monitoring
Stuck and zombie workers are detected automatically. Failed tasks retry up to configurable limits before the convoy marks them as permanently failed.
Mixed Runtimes
Per-task adapter overrides let you run Copilot, Claude Code, Cursor, OpenCode, Windsurf, Codex, and Antigravity workers in the same convoy.
Validation Gates
Run lint, tests, and builds after all tasks complete. Gates must pass before the convoy is marked done — no silent failures.
Generate Convoy
Describe your goal in plain language. The Team Lead
decomposes it into a valid convoy.yml spec
with dependencies, file partitions, and self-contained
agent instructions.
Track every delegation, every session, every lesson.
A real-time dashboard shows agent performance, model tier distribution, session timelines, and panel review outcomes.
Plugins for your entire stack.
OpenCastle ships with 31+ plugins — each bundling skills, MCP server configs, and agent tool mappings for popular tools and services. Choose your stack during onboarding and the CLI wires everything up.
Reproducible execution plans for every task type.
Advanced workflow templates define phases, agent assignments, exit criteria, and file partitions. Handy prompt templates cover common tasks from brainstorming to PR reviews.
| Template | Pipeline |
|---|---|
| feature-implementation | Brainstorm → Research → Foundation → Integration → Validation → QA Gate |
| bug-fix | Triage & Reproduce → Root Cause Analysis → Fix → Verification |
| data-pipeline | Source Analysis → Scraping → Processing → Validation → Import |
| security-audit | Scope → Automated Checks → Manual Review → Panel Review → Remediation |
| performance-optimization | Baseline Measurement → Analysis → Optimization → Verification |
| schema-changes | Schema Analysis → Implementation → Query Updates → Page Integration → Verification |
| database-migration | Planning → Migration → Type Generation → Code Integration → Rollback Test |
| refactoring | Scope & Baseline → Test Coverage → Refactor → Verification → Panel Review |
No code ships without verification.
Multi-layered validation gates — from secret scanning to final smoke tests — with structured dispute escalation when agents can't converge.
01 Secret Scanning
Every diff is scanned for API keys, tokens, passwords, and connection strings before any other gate runs. Leaked secrets block immediately.
02 Deterministic Checks
Lint, type-check, unit tests, and build verification run automatically after every agent output. No false positives.
03 Blast Radius & Dependencies
Scope analysis flags unexpectedly large diffs and sensitive file changes. New npm packages are audited for vulnerabilities, licenses, and bundle size.
04 Fast Review
A single reviewer sub-agent validates every delegation. Sensitive files (auth, migrations, security headers) always get reviewed — even for 1-line changes.
05 Browser Testing
Chrome DevTools MCP validates UI changes at your project-defined responsive breakpoints. Screenshot evidence is mandatory.
06 Panel Majority Vote
Three isolated reviewer sub-agents evaluate high-stakes changes. 2/3 majority wins. BLOCK items become fix requests, not stop signals.
07 Final Smoke Test
After all tasks are done, the complete feature is verified end-to-end — catching integration issues that per-task testing misses.
08 Structured Disputes
When agents exhaust all automated resolution, formal dispute records package both perspectives and resolution options — giving humans clear action items.
A complete orchestration framework, out of the box.
86K+ words of curated knowledge, and everything you need to run a coordinated AI development team — from agent definitions to quality gates.
Support OpenCastle
OpenCastle is free and open-source. If it saves you time, consider buying us a coffee to keep development going.