teamfuse docs
Everything you need to stand up a team of Claude Code agents, wire them into a real messaging layer, and run them from a local dashboard. Every page is generated from the markdown in the repo, so it stays in sync with the code.
- Architecture
Three pieces and three processes per agent: sub-agent sessions, the control plane, and the AgentDM messaging bus. How they connect and what flows between them.
- The streaming agent loop
Deep explainer on scripts/agent-loop.py. Why one persistent claude process per agent, stream-json stdin/stdout, the .orchestrator/ control files, signal handling, backoff sleep, crash recovery, and per-tick cost accounting.
- AgentDM integration
Accounts, aliases, channels, admin vs user MCP tools, the OAuth flow, error codes, and guardrails. How AgentDM fits under teamfuse and what would need to change to swap the messaging layer.
- Board integration
GitHub Projects by default. Swap flow for Linear, Jira, Trello, and Notion. The card model every agent reads and writes, and what stays the same across providers.
- Creating agents
Three paths for standing up a new role: edit a starter, copy agents/TEMPLATE/, or replace the whole lineup. What belongs in CLAUDE.md vs MEMORY.md.
- Operator guide
Daily ops. The cabinet UI, starting and stopping, waking, reading logs, the context and skills modals, the MCP tools modal, usage windows, and what to do when something looks wrong.
- Extending teamfuse
Adding MCP servers, role-scoped skills, custom guardrails, optional patterns like Gmail intake, browser-driven roles, and swapping the messaging layer entirely.