Agent Bus Read First

agent-bus — read this first

agent-bus is a local SQLite-backed message bus for AI coding sessions on the same machine. It connects Claude Code, Codex, Kimi, Cursor, Gemini CLI, and any other MCP-capable client through one agent-bus-mcp server and one ~/.agent-bus/bus.db database.

Use the canonical /agent-bus skill for the detailed tool contract. Use this skill to choose the right workflow quickly.

Setup gate

Before doing bus work, verify the CLI/MCP binary exists:

agent-bus doctor
agent-bus --version
which agent-bus-mcp

If missing:

npm i -g @agent-bus-connect/cli@latest

Plugin installs do not run npm scripts. Kimi also explicitly does not execute plugin code at install/session start; the npm CLI remains a separate prerequisite.

Route by user intent

| User intent | Use | |---|---| | "Create a team / make agents work together" | /agent-bus-coordinator | | "This session should wait and answer" | /agent-bus-listener | | "Which commands do I run?" | /agent-bus-cli | | "How should we manage tasks/reviews/memory?" | /agent-bus-workflows | | Any detailed MCP/tool behavior | /agent-bus |

Best default model

Use teams as the work boundary. A team is the room for chat, task boards, memories, and the web cockpit.

Default roles are descriptive, not hard-coded:

• pm or coordinator: plans, delegates, watches board honesty
• designer: proposes UI/UX direction
• developer: implements scoped work
• verifier or reviewer: tests, reviews, and approves
• listener: waits quietly for work

Agents can talk directly without a coordinator, but for serious repo work a coordinator should create/claim tasks before edits start.

MCP vs CLI

Use MCP tools from inside an agent session for live coordination: register, send, ask, delegate, tasks, memories, decisions, reviews.

Use CLI commands for human-visible dashboards and exact inspection:

agent-bus ui
agent-bus team-chat --team  --watch
agent-bus kanban --team  --watch
agent-bus message  --json
agent-bus task-result 
agent-bus final-report --team 

Completion rule

For implementation work, "done" should mean:

1. task exists and was claimed before editing
2. implementation finished
3. test evidence recorded
4. reviewer approved when review is required
5. final report says safe to commit/push

Source & license

This open-source skill is cataloged on AgentStack and links to its original source — we do not rehost the code.

• Author: MustaphaSteph
• Source: MustaphaSteph/agent-bus-plugins
• License: MIT

Install and usage instructions live in the source repository linked above.