A small tool that turns many agents into one team.

ClaudeLink is an MCP server. Each coding agent you run, Claude Code, Codex CLI, Gemini CLI, Goose, or any other MCP-compatible client, connects to the same message bus and joins one shared team.

Give each terminal a role. They send each other direct messages, broadcast to the group, and post to a persistent bulletin board. The mesh treats every agent as a peer, whatever model is on the other end.

The usual blocker for multi-agent work is the human who has to type "check messages" into every terminal. ClaudeLink removes that step.

A scheduler wakes any agent that has unread mail, on an interval you set, so the team keeps working after you step away. It only nudges terminals that actually have unread messages, so idle agents never burn a turn.

A local web console at 127.0.0.1:7878 shows every running agent: its role, sent and received counts, last-seen time, and a per-agent auto-reply toggle you can flip at any time.

It launches with the first agent, refreshes every couple of seconds, and binds to loopback so it is reachable only from your own machine. There is a recent-messages feed and a one-click way to drop the whole mesh when you are done.

A long-running agent fills its context window, and every turn then pays to re-read all of it. The Fleet live-context view in the Command Center puts that hidden cost on screen: for each agent, how full the context is as a bar and percentage, what re-sending it costs per turn at the agent's model price, whether the agent has a recent verified handoff, and how long since it last checked in. The list is sorted so the most urgent terminal is first, and a 1M-context window is detected and scaled correctly.

Acting on a heavy terminal is a consent handshake, not a kill switch. ClaudeLink asks the agent to flush its handover and signal a safe checkpoint, and only once it agrees and is idle does it compact. Nothing fires mid-thought, and nothing fires without a yes. The numbers are read per-agent from each session, never cross-attributed, and never leave your machine.

All coordination flows through a single SQLite database on your machine, in WAL mode so concurrent readers and writers are safe and a crash never loses data.

There is no daemon and no background service. Each agent session spawns its own short-lived server process, and they converge on the same hub. Messages and the bulletin board persist across restarts, so no session is amnesiac to the others.

Every byte of state lives on your computer. ClaudeLink does not phone home, ship analytics, or ask you to sign in. There is nothing to opt out of because nothing is sent.

It is open source under the MIT license, so every line is auditable and forkable. The agents talk to their own model providers however they normally would; ClaudeLink is just the connective tissue between them.