MCP Integration

Fleet exposes 26 tools via the Model Context Protocol (MCP), enabling Claude Code and other MCP clients to manage your fleet programmatically.

Setup

1. Start the MCP Server

The MCP server communicates over stdio using JSON-RPC 2.0:

fleet mcp serve

2. Configure .mcp.json

Add this to your project root so Claude Code auto-discovers the server:

{
  "mcpServers": {
    "fleet": {
      "command": "fleet",
      "args": ["mcp", "serve"],
      "env": {}
    }
  }
}

How it works

When Claude Code starts, it reads .mcp.json, launches fleet mcp serve as a subprocess, and communicates via JSON-RPC 2.0 over stdin/stdout. All 26 tools appear as available tools in the Claude Code session. No network ports, no HTTP -- just stdio.

Operator Tools

18 tools for fleet management -- used by humans via Claude Code.

Tool	Description
fleet_status	Get fleet health summary with agent statuses
fleet_agent_list	List all configured agents with current status
fleet_agent_start	Start an agent by name
fleet_agent_stop	Stop a running agent
fleet_agent_detail	Get detailed agent information
fleet_agent_output	Read recent output from an agent's tmux session
fleet_agent_budget	Get budget utilization for an agent
fleet_team_list	List all configured teams
fleet_team_detail	Get team details and member agents
fleet_log	Query unified fleet log with filters
fleet_pipeline_list	List all pipeline definitions
fleet_pipeline_run	Start a new pipeline run
fleet_pipeline_status	Get status of a pipeline run
fleet_pipeline_signal	Signal stage completion for a pipeline run
fleet_fabric_events	Query fabric event bus
fleet_fabric_publish	Publish an event to the fabric bus
fleet_task_assign	Assign a task to a specific agent
fleet_eval	Run evaluation scoring on an agent

Agent-Facing Tools

5 tools designed for agents to use during their autonomous runs.

Tool	Description
fleet_inbox	Check the agent's fabric inbox for messages
fleet_inbox_ack	Acknowledge and process an inbox message
fleet_claim	Claim a fabric event for processing
fleet_publish_decision	Publish a structured decision event
fleet_signal_complete	Signal pipeline stage completion

Config Tools

3 tools for reading fleet configuration.

Tool	Description
fleet_config_show	Display current fleet configuration
fleet_config_agents	List agent configurations from config files
fleet_config_pipelines	List pipeline configurations

Conversational Examples

With MCP configured, you can manage your fleet through natural conversation in Claude Code:

"Show me the fleet status"

Uses: fleet_status

Returns color-coded health summary of all agents, watchers, and pipelines.

"Start the frontend-dev agent and assign it the dark mode task"

Uses: fleet_agent_start + fleet_task_assign

Starts the agent in a tmux session, then dispatches the task with FLEET_TASK_TITLE.

"What has the QA team been doing in the last hour?"

Uses: fleet_log

Queries the unified fleet log filtered by team and time range.

"Run the deploy pipeline for the v1.13 release"

Uses: fleet_pipeline_run

Creates a new pipeline run, starts first-stage agents.

"Check if frontend-dev has any inbox messages"

Uses: fleet_inbox

Returns unread fabric inbox messages for the agent.

Protocol Details

Transport

JSON-RPC 2.0 over stdio (stdin/stdout). No HTTP server, no WebSocket, no network ports. The MCP client spawns fleet mcp serve as a child process and communicates via pipes.

Request Format

{
  "jsonrpc": "2.0",
  "id": 1,
  "method": "tools/call",
  "params": {
    "name": "fleet_agent_start",
    "arguments": {
      "name": "frontend-dev"
    }
  }
}

Response Format

{
  "jsonrpc": "2.0",
  "id": 1,
  "result": {
    "content": [
      {
        "type": "text",
        "text": "Agent frontend-dev started successfully"
      }
    ]
  }
}

Discovery

The MCP server responds to tools/list with all 26 tool definitions including parameter schemas. Claude Code calls this automatically on startup to populate its tool list.