agentchan MCP server

agentchan

An MCP server that lets one coding agent talk to another: Claude Code, Codex or Grok. It gives one tool for each agent that is installed and signed in, plus two tools that work like Go channels:

once. Without channel, the message starts a new session. With channel, it is a follow-up.

• claude(message, channel?), codex(…), grok(…) send a message and return a channel id at

• receive(channels) waits for the first reply on any of the channels, up to 50 s.
• close(channel) stops the agent and ends the channel.

Each channel is one agent session. Each message runs claude -p, codex exec or grok --prompt-file in the server's working directory, with permission checks off, and with the settings that agent already has.

Install

Linux and macOS, prebuilt:

curl --proto '=https' --tlsv1.2 -LsSf https://github.com/damejeras/agentchan/releases/latest/download/agentchan-installer.sh | sh

From source:

cargo install agentchan

The server offers every agent that is installed and signed in when it starts.

# Claude Code
claude mcp add --scope user agentchan -- agentchan

# Codex: ~/.codex/config.toml
[mcp_servers.agentchan]
command = "agentchan"

Security

The called agent runs with permission checks off. It can read, write and run anything in the server's working directory without asking. Any agent that can call agentchan gets that access.

When a turn ends, and when a channel closes or the server stops, agentchan kills the agent's process group. A process that the agent moved to a different process group keeps running.

Test

cargo test                 # fake agents, no plan limits used
cargo test -- --ignored    # the real CLIs that are signed in

Cost

Each message uses the called agent's own login and plan limits. Each message starts a new process that loads that agent's full configuration again.

Links

• MCP Registry name: mcp-name: io.github.damejeras/agentchan