AI编程代理共享收件箱

Agent Bus MCP

Agent Bus MCP is a shared local inbox for AI coding agents.

Use it when you want Codex, Claude Code, Gemini CLI, OpenCode, Cursor, or another MCP-capable tool to work on the same task without copy-pasting summaries, losing context after restarts, or scattering coordination across scratch files.

Under the hood, Agent Bus MCP is a stdio MCP server backed by SQLite. Agents create named topics, join with stable peer names, send and receive messages through sync(), and resume from server-side cursors. The result is a durable, searchable record of handoffs, reviews, and sidecar work that stays on the local machine.

[!TIP] Prefer the polished product and docs front door? Visit agentbusmcp.com for the live website, then use this README and docs/ when you want the repo-native setup, reference, and implementation details.

Why use Agent Bus MCP?

Multi-agent coding breaks down when every tool keeps its own memory. One agent implements, another reviews, a sidecar investigates, and the coordination ends up in pasted summaries, terminal logs, or repo scratch files.

Agent Bus MCP gives those agents one shared place to coordinate:

• Handoffs stay readable. Each task gets a named topic with one ordered message stream.
• Agents keep their identity. A reviewer, implementer, or sidecar can reconnect with the same peer name.
• Restarts do not erase progress. Server-side cursors remember what each agent has already seen.
• History stays inspectable. Use the CLI or Web UI to replay, export, and search past coordination.
• Everything stays local. The bus runs over stdio and stores state in SQLite. No hosted service is required.

When Agent Bus MCP is a good fit

Agent Bus MCP is a strong fit when:

• two or more local coding agents need to collaborate on the same task
• an agent should be able to disconnect and later reclaim its identity
• you want durable replay instead of "read everything again" polling
• you want a searchable local log of agent-to-agent coordination

It is a weaker fit when:

• you only need a single agent with no handoffs
• you need networked multi-host messaging
• you need auth, tenancy, or cloud-hosted coordination out of the box

Quickstart

The fastest path is to install Agent Bus MCP as an MCP server in your client and start using it from natural-language prompts.

export AGENT_BUS_VERSION="0.5.1"
npx install-mcp "uvx --from agent-bus-mcp==$AGENT_BUS_VERSION agent-bus" --name agent-bus --client claude-code

Replace 0.5.1 if you want a different release. For direct setup:

# Codex
codex mcp add agent-bus -- uvx --from "agent-bus-mcp==$AGENT_BUS_VERSION" agent-bus

# Claude Code
claude mcp add agent-bus -- uvx --from "agent-bus-mcp==$AGENT_BUS_VERSION" agent-bus

Then ask an agent to:

1. create or reuse a topic
2. join the topic with a stable agent_name
3. send or read messages through sync()

For a fuller walkthrough, start with First topic between two peers.

Web UI

Agent Bus MCP also ships with a local Web UI for browsing topics, reading threads, exporting logs, and searching the bus without leaving the browser.

From a source checkout:

uv sync --extra web
pnpm --dir frontend install
pnpm --dir frontend build
uv run agent-bus serve

Or launch the published package directly with the web extra:

uvx --from "agent-bus-mcp[web]==$AGENT_BUS_VERSION" agent-bus serve

The overview keeps recent topics in the sidebar and leaves the main workbench focused on navigation and search.

For the full walkthrough, including thread navigation, export, and troubleshooting, see How to use the Agent Bus MCP Web UI.

Documentation

This repo now has a lightweight Diataxis-style docs layout under docs/:

Recommended starting points:

• Why use Agent Bus MCP?
• Install and configure Agent Bus MCP
• Runtime reference
• Implementation spec

Upgrading from an older release? See CHANGELOG.md.

Optional workflow skill

This repo also includes a reusable workflow skill asset at ./.agents/skills/agent-bus-workflows/.

It is useful when you want ready-made reviewer/implementer loops, handoffs, duplicate-name recovery, and reclaim-token reconnect behavior in a Codex-style skill package. See Install and configure Agent Bus MCP for the install path.

Architecture at a glance

%%{init: {"look": "handDrawn", "fontFamily": "virgil, excalifont, segoe print, bradley hand, chalkboard se, marker felt, comic sans ms, cursive", "flowchart": {"diagramPadding": 130}, "htmlLabels": true}}%%
flowchart TB
  subgraph Clients
    A1[Agent MCP client]
    A2[Agent MCP client]
  end

  subgraph Python
    BUS[agent-bus MCP server over stdio]
    CLI[CLI]
    WEB[Web UI]
    IDX[Embeddings indexer]
  end

  CORE[(Rust core / agent-bus-core)]
  DB[(SQLite DB)]

  A1 <--> BUS
  A2 <--> BUS
  BUS --> CORE
  CLI --> CORE
  WEB --> CORE
  IDX --> CORE
  CORE --> DB

The Python package provides the MCP server, CLI, Web UI, and embedding worker. The Rust core owns the SQLite schema, reads/writes, search, and embedding-job coordination, which keeps the data model single-sourced.

For the rationale behind this shape, see Why use Agent Bus MCP?.