Cerebro MCP is a universal AI orchestrator that makes Claude Chat the "brain" and CLI workers the "hands." You talk, it builds. 28 MCP tools, multi-provider CLI (Claude Code + OpenAI Codex), agent swarms you create by talking, and the first project to unify both MCP and A2A protocols. Open source, Apache 2.0.
GitHub: github.com/Synvoya/cerebro-mcp
The Problem
There's a disconnect in AI-assisted development that nobody's solving.
Claude Chat is brilliant at planning, brainstorming, and understanding context — but it can't write files or run builds. Claude Code CLI can build software — but you need to know your way around a terminal. Existing orchestrators like Ruflo (259 tools) and Gas Town (60+ agents) are powerful but explicitly target developers.
What about the 99% of people who think in ideas, not code? Product managers, designers, entrepreneurs — people who can describe exactly what they want but can't npm install their way to a working product.
The Solution: Cerebro MCP
Cerebro MCP is an MCP server that turns Claude Chat into an autonomous orchestrator. You describe what you want in plain English. Cerebro decomposes the task, routes it to specialist agents, spawns CLI workers in visible Terminal windows, and reports back in language you understand.
You never touch a terminal. You never write code. You just talk.
Here's what a real session looks like:
You: "Build me a landing page for my coffee shop"
Cerebro: Breaking this into tasks...
→ Designer agent (Claude Code): Creating layout with hero, menu, CTA
→ Coder agent (Claude Code): Building responsive HTML/CSS
→ QA agent (Codex): All 12 tests passing. Lighthouse: 96.
Your landing page is ready at localhost:3000.
Each agent runs in its own branded Terminal window. You can literally watch the AI working.
What Makes It Different
Multi-provider CLI. Cerebro isn't locked to one AI. In the demo above, the Designer and Coder use Claude Code while QA uses OpenAI Codex. You configure this with natural language:
You: "Use Codex for all QA and code review tasks"
Cerebro: Routing updated. QA tasks → Codex CLI.
Cerebro auto-detects installed providers and intelligently routes based on agent name, description, and your rules.
Dual protocol: MCP + A2A. Cerebro is the first orchestrator to implement both Linux Foundation standards. MCP handles tool communication (how Cerebro talks to CLI workers, databases, GitHub). A2A handles agent-to-agent communication (how Designer delegates to Coder, Coder delegates to QA). This means any A2A-compliant external agent can join your swarm.
Agent swarms via conversation. No YAML. No config files. You create agents by describing them:
You: "Add a marketing agent that writes casual social media posts"
Cerebro: Marketing agent created and added to your swarm.
Agents persist across sessions, learn your preferences, and delegate to each other.
Fresh context per task. Every agent call gets a clean context window. Definitions and preferences persist, but working context is wiped after each task. No stale reasoning accumulates. This is how you prevent the context rot that plagues long-running AI sessions.
Session tokens. When your chat hits context limits, Cerebro generates a cryptographically signed handover token. Paste it in a new chat and continue exactly where you left off.
Architecture
User (plain English)
↓
Claude Chat (brain — plans and delegates)
↓
Cerebro MCP Server (28 tools, orchestrates everything)
├── MCP bus → CLI workers (Claude Code, Codex, Aider)
├── A2A bus → Agent swarm (Designer, Coder, QA, Marketing...)
└── External MCPs (GitHub, Slack, databases)
↓
Agent Marketplace (Anthropic Skills + Community + User-created)
↓
Working Product
The Tech
69 files, ~6,500 lines of TypeScript
28 MCP tools across 6 categories (Session, Tasks, Agents, Workers, Vision, Handover)
40/40 tests passing, builds clean
SQLite for persistence, HMAC-SHA256 for token signing
Visible Terminal windows via AppleScript (you watch agents work)
Auto-opens HTML files in browser after task completion
Try It
bashnpx @synvoya/cerebro-mcp
Add to Claude Desktop config and restart. All 28 tools appear in Chat.
What's Next
This is v0.1.0. On the roadmap: Cowork integration (pending Anthropic API), remote CLI workers on VPS, agent dashboard web UI, and a community agent repository.
If you're interested in contributing or just want to try it out, the repo is open: github.com/Synvoya/cerebro-mcp
Top comments (0)