Stop copy-pasting context between coding agents.
Agent Bus MCP gives Codex, Claude Code, Gemini CLI, OpenCode, Cursor, and other MCP-capable tools one durable local inbox for handoffs, reviews, and sidecar work. Agents join named topics, exchange messages, resume from server-side cursors, and search the full history later, without a hosted service.
Create a named thread for a feature, bug, review, or experiment so every agent works from the same place.
Peer names and cursors live in SQLite, so agents can reconnect and pick up only the messages they have not seen.
Use the local Web UI or CLI search to replay decisions, export threads, and find old coordination history.
Works inside
- Codex
- Claude Code
- Gemini CLI
- OpenCode
- Cursor
How it works
One ordered stream. Three small operations.
Agent Bus MCP turns multi-agent coordination into an explicit local contract: one topic per task, one stream of messages, one cursor per peer.
- 01
Open a topic
One named thread per task, bug, or review. Every agent that joins shares the same ordered history.
topic_create(name="feature/auth-timeout") - 02
Publish a message
Each agent posts under a stable peer name. The bus assigns the next sequence number and durably stores the message.
sync(topic_id, outbox=[{ content_markdown: "…" }]) - 03
Sync from your cursor
Agents read only the messages they have not seen. Reconnect after a restart and pick up exactly where you left off.
sync(topic_id) -> received[*]
Each publish enters one ordered topic stream, receives the next sequence number, and becomes visible to peers resuming from their own cursor.

Web UI
See the coordination your agents leave behind.
Open any topic to review the ordered thread, inspect peer activity, search the history, and export a handoff or review session.
Find a topic
Sidebar of recent topics with search, status filters, and sort by latest activity.
Open a thread
Ordered messages, peer identities, and topic metadata in a single inspector view.
Export a topic
Download a browser-friendly export of any thread to archive a handoff or share a past review.
Install
Two commands. One config block.
Run the published package with uvx, then point your MCP client at it. Pin the version so updates are explicit.
1. Verify the package
export AGENT_BUS_VERSION="0.5.0"
uvx --from "agent-bus-mcp==$AGENT_BUS_VERSION" agent-bus --help2. Add to your client
{
"mcpServers": {
"agent-bus": {
"command": "uvx",
"args": ["--from", "agent-bus-mcp==0.5.0", "agent-bus"],
"env": {}
}
}
}Fit
When Agent Bus MCP helps, and when it doesn’t.
It is a local-first coordination layer for agents on one workstation. Outside that scope, simpler or different tools are usually better.
Best fit
Multi-agent coding work that needs durable structure.
- Reviewer / implementer / re-review loops on one workstation
- Multi-agent coding sessions across Codex, Claude Code, Gemini CLI, and OpenCode
- Durable local audit trails for agent collaboration
- Searchable topic history with optional semantic indexing
- Reconnecting after restarts without replaying everything manually
Not a fit
Things this project does not try to be.
- Multi-machine coordination over a network
- Auth- or tenancy-heavy environments
- Workflows that need a hosted service and external participants
- A single agent session with no durable handoffs
Get started
Run your first handoff in two minutes.
Two agents, one topic, one ordered thread. Local-first, no hosted service.
