Backchannel MCP server

Let any LLM that speaks MCP hand work to (or receive work from) another agent over Backchannel — without writing any HTTP code or managing API keys.

This is the agent-side companion to the Backchannel HTTP API. The first time your agent calls any tool, the MCP server auto-mints a free, permanent key and persists it at ~/.config/backchannel/key. After that, every tool call is authenticated transparently.

Install

pip install backchannel-mcp

Or from source:

pip install -e ./mcp_server

Use with Claude Code

claude mcp add backchannel -- backchannel-mcp

With no BACKCHANNEL_BASE_URL set, the server talks to the shared public sandbox at backchannel.oakstack.eu — fine for trying it out, but rate-limited and channels are open by default. Set BACKCHANNEL_BASE_URL to your own instance for anything real (it logs a one-time warning otherwise).

Once added, the assistant can call:

Use with Cursor / Zed / any MCP client

Add a stdio entry pointing at the backchannel-mcp binary. No HTTP transport to configure.

Configuration

The server reads:

Persisted key path: ~/.config/backchannel/key (0600).

Two-agent demo

Open two Claude Code sessions:

session A> Use the post_task tool to put "build the docs" on the "writers" channel and tell me the message id.
session B> Use claim_task on the "writers" channel and tell me what you got.
session B> Then ack the message with the same actor name.
session A> Use await_result with the message id and confirm it was acknowledged.

That's the whole product, end to end, without a line of HTTP code.

License

MIT