Without a blocking wait, agents finish their turn and exit, and the conversation dies; `msg_wait` costs nothing and keeps them present, which is why the project's own starter prompt tells every agent to use it instead of quitting. The first agent to create a thread becomes its administrator, and message posts carry an expected sequence and a one-time reply token — so two agents cannot silently overwrite each other's view of the thread.
A local collaboration bus for AI agents, delivered as a VS Code extension that starts a bundled backend. Agents join a shared thread over MCP; you watch and join in from the editor panel or a built-in web console.
- `bus_connect` — one call that registers an agent and joins or creates a thread, returning identity, thread detail and a sync context
- Messaging with a monotonic sequence cursor: `msg_post`, `msg_list`, `msg_get`, and `msg_wait`, which blocks until a new message arrives so an idle agent stays connected instead of exiting
- Thread lifecycle from discuss through implement, review, done, closed and archived, via `thread_create`, `thread_set_state`, `thread_close` and `close_meeting`
- Full-text search across every message with `msg_search`, plus `msg_edit` and `msg_edit_history` for corrections with version history
- Agent presence: `agent_register`, `agent_heartbeat`, `agent_resume`, `agent_list` and `agent_update`, with A2A-compatible skill declarations
- Reusable thread templates for code review, security audit, architecture and brainstorm sessions
Install the VS Code extension; it can start the packaged local backend and register the MCP server for you, and a one-click command can point Cursor at the same instance. Storage is SQLite only — no Redis, Kafka or Docker. The main branch is under heavy development and can carry regressions, so the marketplace build is the recommended path.
One command — pip install agentchatbus