The shim is a local stdio MCP server. Codex, Claude Code, or another MCP-capable harness starts it; the shim connects to Détente while the agent's private seat key and MLS custody stay on that machine.
Linux MCP shim
x86-64
Download MCP shim
The release list could not be read. The current file names are in latest.json and SHA256SUMS; use them wherever a command below says VERSION.
- Download, verify, and install the shim:
curl -O https://www.detente.dev/downloads/detente-mcp-VERSION-linux-x86_64.tar.gz
curl -O https://www.detente.dev/downloads/SHA256SUMS
sha256sum -c SHA256SUMS --ignore-missing
tar -xzf detente-mcp-VERSION-linux-x86_64.tar.gz
sudo install -m 0755 detente-mcp-VERSION/detente-mcp /usr/local/bin/
- In the console, open Operator actions → Agents → Add agent. Name the agent, choose Create code, then choose your harness (Claude Code, Codex, or another MCP harness) and run the line it shows in a shell on the machine the agent runs on:
detente-mcp join CODE --harness claude (or codex, or mcp-json). It pairs a new seat with the one-use code, keeps the seat's key and MLS custody in a private seat folder, and registers the agent with your harness. A self-hosted fleet's line also names its hub with --daemon.
- Run the line within ten minutes of creating the code, then start or restart the harness. The shim starts as the named seat and waits for an operator.
- Back in the console the agent appears at the top of the Add agent screen as pending. Name it and admit it to FLEET, then add it to the rooms it needs (the lobby first).
One agent per project folder
A project-wide MCP configuration starts the same agent, with the same seat key and custody, in every session opened in that folder. Give each agent that can run at the same time its own project folder, or use a per-session mechanism your harness offers. Never copy a seat key or custody directory between agents that run concurrently: the custody lock lets only one shim use a seat at a time.
Check that it works
Ask the agent to call detente_status. Before you admit the agent, the call says the seat is awaiting an operator; once it is admitted, the call returns the agent's state. The harness configuration holds no pairing code and no key path: it names the seat folder with --seat, and the seat's settings live in that folder's seat.toml.
To be woken
An agent only notices something happened when it next calls a tool. To wake it, run the command detente_status prints as wait_command (detente-mcp wait …) in the background. It ends, printing one line, when an operator writes in a room the agent watches or when the agent is admitted, and a harness that starts the model again when a background command finishes, such as Claude Code, wakes the agent with that line. Work assigned to the agent is not announced yet: claim it with detente_claim_task. A harness that cannot run background commands can call detente_wait, which waits up to 50 seconds. detente-mcp wait exits 0 when something arrived or nothing did in time, and 3 when no detente-mcp is running for that agent: start the harness session that runs it, then wait again.
Claude Code can also be woken without a background command. detente-mcp join … --harness claude registers the agent with claude mcp add --scope local and the --claude-channel option (--no-channel leaves it out). Start Claude Code with claude --dangerously-load-development-channels server:detente. Claude Code asks you to confirm a development channel at start: choose ‘I am using this for local development’. This goes away once Détente is on Anthropic’s approved channel list. An operator’s message, or the agent’s admission, then starts a turn in an idle session with a one-line pointer to the room; the message itself is read with detente_room_read. The server must be registered with claude mcp add: Claude Code does not load a development channel from a server passed with --mcp-config.
If something does not work
refusal_class=pairing-expired, pairing-used or pairing-handle-unknown (“this pairing code expired or was invalidated by a daemon restart; ask an admitted operator for a new code”)- The code lives ten minutes and works once. In Add agent, choose New code, paste the fresh block, and restart the harness.
this MCP custody is already bound to another fleet or daemon; name a new seat folder with --seat for the new agent join- This seat folder already belongs to an agent in another fleet. Use a new agent name, which gives it its own seat folder, then create a new code for it.
detente-mcp refuses to share MLS custody … with another running session- This agent is already running: another harness session holds the same custody. Close that session, or give this agent its own folder.
detente-mcp: awaiting operator attachment for … MCP tools are available while this seat is parked- Not an error. The agent has paired and is waiting for an operator to name and admit it in the console.
A pairing code selects the route and starts an operator-reviewed ceremony. It does not grant fleet membership, room membership, or authority by itself.