The governance layer is the reason to run it: assistant tool calls, integration writes and git pushes go through a risk policy that auto-runs, asks, or rejects by tier, with an audit trail. The tool count is large, so the compact /mcp endpoint folds 268+ tools into roughly 33 meta-tools for clients like Claude.ai that cap tool counts, while /mcp/full exposes them individually.
FleetQ — a self-hosted platform for building, running and monitoring multi-agent systems, with a visual DAG builder, human-in-the-loop approvals and an MCP server that exposes the platform to any LLM.
- Visual DAG workflows with 8 node types — agent, conditional, human-task, switch, dynamic-fork, do-while, compensation and sub-workflow
- Multi-agent crews with coordinator, worker and reviewer roles and weighted QA scoring
- Real-World Action governance — tool calls, integration writes and git pushes route through a per-tier risk policy of auto, ask or reject
- A credit ledger with pessimistic locking and auto-pause on overspend
- The full server exposes 268+ MCP tools; the compact endpoint at /mcp consolidates them into about 33 meta-tools for clients with tool-count limits
- Structured MCP errors with gRPC-canonical codes and retry hints, and an optional deadline_ms per tool call
- tool_ssh_fingerprints and tool_bash_policy manage SSH trust and bash risk policy from MCP
A self-hosted deployment — a Laravel 13 monolith, run via Docker Compose. The MCP server is reachable over HTTP/SSE with a Sanctum bearer token (or OAuth2 in cloud), or over stdio with php artisan mcp:start agent-fleet for a local CLI agent. Get a token from the API or the Team Settings UI. AGPLv3. A managed FleetQ Cloud tier exists if you would rather not self-host.
Build from source — clone the repository and build it