Communication hub for agents — multiple Claude Code instances collaborate via Unix Domain Sockets.
Claude Code's subagent (Task tool) is a "delegation model":
- Humans can't see intermediate steps
- Agents can't communicate directly with each other
- No control over agent lifecycle
AgentMesh solves these problems by running each Claude Code in its own tmux pseudo-terminal:
- Humans interact with AI directly through the terminal (normal Claude Code usage)
- AIs communicate with each other via Unix Domain Socket + MCP
- Humans can switch terminals at any time to observe or intervene with any agent
┌──────────────┐ ┌──────────────┐
│ tmux pane A │ │ tmux pane B │
│ │ │ │
│ Human ↔ AI │ │ Human ↔ AI │
│ │ │ │
│ MCP Server │ │ MCP Server │
│ ↓ write │ │ ↓ write │
└──┬───────────┘ └──┬───────────┘
│ │
│ ┌─────────────────────────────┐ │
└──→ ~/agentmesh/ ←──────┘
│ │
│ registry.json │
│ sock/ │
│ agent-<pid-1>.sock │
│ agent-<pid-2>.sock │
└─────────────────────────────┘
- Platform: macOS (Linux support planned)
- AI: Claude Code
- Runtime: Python >= 3.12, tmux >= 3.3
git clone https://github.com/akachi10/agentmesh.git
cd agentmesh
bash install.sh# Start an agent
amesh
# Open another terminal and start a second agent
ameshEnter a name (e.g. "architect", "developer") and optional custom instructions, then you're in a Claude Code session.
AI automatically communicates with other agents via MCP tools list_agents() and send_message().
bash uninstall.sh| Component | Choice |
|---|---|
| Language | Python 3.12+ |
| IPC | Unix Domain Socket |
| Terminal | tmux >= 3.3 |
| AI Tool Protocol | MCP (stdio) |
| Registry | JSON + flock |
MIT