Skip to main content

Claude Code

fleet-bridge is a stdio MCP server, so Claude Code launches it as a subprocess. Transport is stdio only — there is no HTTP transport in v0.1.

Server registration​

Add the server to your MCP config (command + args, optional env):

{
"mcpServers": {
"fleet-bridge": {
"command": "node",
"args": ["/absolute/path/to/fleet-bridge/dist/index.js"]
}
}
}

If the rosters live outside the repo, or you want to override timeouts, set them on the server entry:

{
"mcpServers": {
"fleet-bridge": {
"command": "node",
"args": ["/absolute/path/to/fleet-bridge/dist/index.js"],
"env": {
"FLEET_BRIDGE_AGENTS": "/absolute/path/to/agents.json",
"FLEET_BRIDGE_HERMES": "/absolute/path/to/hermes-agents.json",
"FLEET_BRIDGE_INVOKE_TIMEOUT_MS": "60000"
}
}
}
}

If you installed the package (rather than running it from a checkout), the fleet-bridge bin entry works too — keep npm run build current, since bin/fleet-bridge.js just imports dist/index.js:

{
"mcpServers": {
"fleet-bridge": {
"command": "npx",
"args": ["-y", "fleet-bridge"]
}
}
}

Using the tools​

After the server starts, Claude Code sees six tools: list_agents, agent_health, start_session, invoke_agent, list_hermes_agents, invoke_hermes. See the tool catalog for per-tool reference.

Notes specific to driving the bridge from Claude Code:

  • Timeouts: Claude Code applies its own timeout to MCP tool calls. If your agents routinely take longer than the client's budget, a longer timeout_ms on the call (or a raised FLEET_BRIDGE_INVOKE_TIMEOUT_MS) will not help unless the client's own ceiling is raised too — the client cancels the call first.
  • Progress: Claude Code supports MCP progress notifications, so long invoke_agent calls emit live character-count updates when it passes a progress token.
  • Restart on roster change: restarting the Claude Code session restarts the bridge and reloads rosters.

Smoke test​

Before wiring the client, verify the server standalone — see Getting started.