crewswarm

A multi-engine AI coding platform. Switch between Claude Code, Cursor, Gemini, Codex, OpenCode, and crew-cli mid-conversation. Parallel agents. Persistent sessions. No vendor lock-in.

Why crewswarm?

Rate limits are real. Every $20/month AI coding plan has them. Claude, Cursor, Codex — you'll hit the wall mid-feature.

crewswarm lets you switch engines without throwing away your working context. Or pick the best CLI for each job:

| Engine | Best for | Key strength | |---|---|---| | Claude Code | Large refactors, frontend | Full workspace context, session resume | | Cursor CLI | Architecture, complex reasoning | Parallel waves, isolated contexts | | Gemini CLI | Research, SEO, free fallback | Free: 60 req/min, Google Search built in | | Codex CLI | Backend, fast iteration | Full sandbox, no approval prompts | | OpenCode | Provider flexibility | Any model (Groq/DeepSeek/Ollama) | | crew-cli | Execution engine, any model | 40+ models, 16 providers, sandbox, parallel workers |

crew-cli: the execution engine

crew-cli is the built-in execution engine. It runs standalone or as one of crewswarm's 6 engines. Unlike other AI coding CLIs that lock you to one provider, crew-cli works with any model:

| Provider | Models | Driver | |---|---|---| | OpenAI | GPT-5.4, GPT-5.2, GPT-4.1, o3-mini, o4-mini | OpenAI API | | Anthropic | Claude Opus/Sonnet/Haiku (API + OAuth) | Anthropic API + CCH | | Google | Gemini 2.5 Flash, Gemini 3 Pro | Gemini API | | xAI | Grok 4.20, Grok 4-1, Grok 3, grok-code-fast | OpenAI-compatible | | DeepSeek | DeepSeek Chat, DeepSeek Reasoner | OpenAI-compatible | | Groq | Llama 4 Scout, Kimi K2, Qwen3-32B, GPT-OSS | OpenAI-compatible | | Fireworks | DeepSeek v3, GLM-5, Kimi K2.5, GPT-OSS | OpenAI-compatible | | Mistral | Mistral Large, Codestral, Devstral | OpenAI-compatible | | Cerebras | Qwen 3-235B, Llama 3.1 | OpenAI-compatible | | NVIDIA | Llama 3.3-70B | OpenAI-compatible | | OpenCode/Zen | 39 models (GPT-5.x, Claude, Gemini, GLM, Kimi, MiniMax, Qwen, Nemotron) | OpenAI-compatible | | OpenRouter | 100+ models | OpenAI-compatible | | Ollama | Any local model | OpenAI-compatible |

Three API formats, one engine. Bring your own keys, use OAuth from Claude Code / Codex / Gemini CLI, or run free models via Groq / Ollama / OpenRouter.

Benchmark: free models match Claude Opus

Single-model pipeline testing — same model handles routing, planning, execution, review, and fixing. All tests pass on 6 TypeScript coding tasks (bugfix, feature, refactor, multi-file).

| Model | Provider | Result | ~Cost/Task | |---|---|---|---| | Qwen 3.5 (397B) | Ollama cloud | All tests pass | FREE | | GLM-5.1 | Ollama cloud | All tests pass | FREE | | Claude Sonnet 4.6 | Anthropic OAuth | All tests pass | ~$0.06 | | Claude Opus 4.6 | Anthropic OAuth | All tests pass | ~$0.07 | | GPT-5.4 | OpenAI OAuth | All tests pass | ~$0.13 |

Two free local models (Qwen 3.5, GLM-5.1 via Ollama) pass every test that Claude Opus does. Full results →

What makes crew-cli different from other AI coding CLIs:

• Failure memory — blocks repeated bad moves, forces new approaches
• Verification-first — won't declare success until proof is shown
• Patch critic — evaluates every edit for quality in real time
• Parallel workers — splits multi-file tasks into concurrent work units
• Sandbox preview — all changes staged before writing to disk
• Smart delegation — picks the right model tier for each subtask

Quickstart

npm install -g crewswarm
crewswarm

That's the default path for most users. Dashboard opens at localhost:4319, Vibe IDE at localhost:3333.

Contributor setup from source

git clone https://github.com/crewswarm/crewswarm
cd crewswarm
bash install.sh
npm run doctor
npm run restart-all

Use source install when you want to work on crewswarm itself or debug local internals.

Docker for servers and teams

curl -fsSL https://raw.githubusercontent.com/crewswarm/crewswarm/main/scripts/install-docker.sh | bash

Use Docker when you want stronger isolation, easier restarts, or a shared server/team box.

What does it cost?

$0. crewswarm is free and open source (MIT). You bring your own API keys — or use CLI OAuth (Claude, Cursor, Gemini login once, no keys needed). Free options: Gemini CLI (1,000 req/day), Groq (free tier), Ollama (fully local).

What makes it different

| Capability | crewswarm | Cursor | Windsurf | Devin | Copilot | |---|---|---|---|---|---| | Multi-engine (6 CLIs) | Yes | No | No | No | No | | Native session resume | Yes | No | No | No | No | | Parallel agent waves | Yes | No | No | Partial | No | | Browser IDE + terminal | Vibe | Desktop | Desktop | Yes | Yes | | 20+ specialist agents | Yes | 1 | 1 | 1 | 1 | | PM Loop (autonomous roadmap) | Yes | No | No | Partial | No | | Swarm Chat (@mention dispatch) | Yes | No | No | No | No | | Local-first / no cloud | Yes | Partial | No | No | No | | Open source | Yes | No | No | No | No |

How it works

1. You write a requirement — one sentence, one paragraph, or a full spec
2. crew-pm plans it — breaks work into phases, assigns specialists
3. Agents execute in parallel — backend, frontend, and tests can run simultaneously
4. Done. Files on disk. — real files, real tests, real output

Dashboard / Vibe IDE / crew-cli / crewchat / Telegram / WhatsApp / OpenClaw / MCP
                    |
                crew-lead (router)
                    |
                 RT Bus
                    |
     ─────────────────────────────────
     |        |        |       |       |
   crew-pm  coder     qa    fixer   github
                    |
        Code Engines: Claude · Cursor · Gemini · Codex · OpenCode · crew-cli
                    |
            real files, commands, memory

Surfaces

• Dashboard — web control plane at localhost:4319 (agents, engines, models, build, sessions)
• Vibe IDE — browser-based editor + terminal + chat at localhost:3333
• crew-cli — execution engine: 40+ models, 16 providers, 41 built-in tools
• crewchat — conversational interface for quick tasks
• Telegram — chat with your crew from your phone
• WhatsApp — mobile messaging bridge
• OpenClaw plugin — use crewswarm as the engineering backend for OpenClaw desktop apps (npm i crewswarm-openclaw-plugin)
• MCP server — plug crewswarm into any MCP-compatible editor

Per-agent model configuration

Every agent gets its own model. Use cheap models for routing, expensive for coding:

{
  "agents": [
    { "id": "crew-lead", "model": "groq/llama-3.3-70b-versatile" },
    { "id": "crew-pm", "model": "google/gemini-2.5-flash" },
    { "id": "crew-coder", "model": "anthropic/claude-sonnet-4-20250514" },
    { "id": "crew-qa", "model": "google/gemini-2.5-flash" },
    { "id": "crew-fixer", "model": "openai/codex-mini-latest" }
  ]
}

Or skip API keys entirely — use Claude Code, Cursor, or Gemini CLI with OAuth login.

The crew

| Agent | Role | |---|---| | crew-lead | Routes tasks, manages conversation | | crew-pm | Plans, breaks down, prioritizes | | crew-coder | Writes code (full-stack) | | crew-coder-back | Backend specialist | | crew-coder-front | Frontend specialist | | crew-qa | Tests and validates | | crew-fixer | Debugs and repairs | | crew-security | Security review | | crew-github | Git, PRs, branches | | crew-architect | System design | | crew-orchestrator | Wave dispatch | | crew-copywriter | Docs and content | | crew-frontend | UI/UX polish | | crew-main | General coordination |

Built with crewswarm

• VS Code extension — full extension from prompt to package in 10 minutes
• crewswarm.ai — the production website you see, 90% built by the swarm in 30 minutes
• Session resume — native resume across 6 CLI engines, built in one session

Commands

crewswarm                    # Start full stack: RT bus, crew-lead, bridges, then dashboard
crewswarm pm-loop            # Run autonomous PM loop
npm run doctor               # Preflight check
npm run restart-all          # Restart the stack
npm test                     # Run the core test suites
npm run test:report          # View test results summary
crew exec "Build X"          # Send task via CLI

Deployment

# Most users
npm install -g crewswarm
crewswarm

# Servers / teams
curl -fsSL https://raw.githubusercontent.com/crewswarm/crewswarm/main/scripts/install-docker.sh | bash

See deploy.html for Docker, cloud VM, and production setup.

Documentation

• Website
• Vibe IDE
• crew-cli
• Models & Providers
• Security
• API Docs
• @@Protocol

Contributing

See CONTRIBUTING.md for guidelines. Issues and PRs welcome.

Security

See SECURITY.md for reporting vulnerabilities.

License

MIT