Agent Fleet O

What escapeboy/agent-fleet-o MCP server does

FleetQ provides a self-hosted control plane for designing, running, and monitoring AI agent systems. The escapeboy/agent-fleet-o MCP server exposes the platform's functions through the Model Context Protocol, allowing compatible LLM clients and agents to query or mutate FleetQ programmatically.

The platform organizes work around agents, skills, experiments, crews, workflows, projects, signals, and MCP tools. Agents can have roles, goals, backstories, skills, provider and model settings, and fallback chains. Crews combine agents in coordinated processes such as sequential, parallel, hierarchical, adversarial, fanout, or chat-room execution. Experiments provide a stateful lifecycle for more involved tasks, while projects can contain one-shot or scheduled work.

How it works

Workflows are built as visual directed acyclic graphs rather than application code. Supported node types include agents, conditions, human tasks, switches, dynamic forks, do-while loops, compensation steps, and sub-workflows. This lets a process branch, repeat, pause for approval, or delegate to another workflow.

Actions can pass through risk policies that classify low-, medium-, and high-risk operations as automatic, approval-required, or rejected. Approved actions can execute automatically, and the platform keeps an audit trail. Budget controls use a credit ledger and can pause work when spending exceeds the configured limit. The platform also provides live execution views, task outputs, metrics, and error details.

The escapeboy/agent-fleet-o MCP server can connect agents to providers including Anthropic, OpenAI, Google, Ollama, LM Studio, vLLM, Codex, Claude Code, and OpenAI-compatible LLMs, according to the project documentation. Local model support and bring-your-own-provider configuration are available alongside hosted providers.

Setup and configuration

FleetQ is intended to run self-hosted with Docker Compose. The project identifies PHP 8.4 and Laravel 13 in its technology stack. It also offers a managed FleetQ Cloud option with a free tier, but that is separate from running the repository yourself.

Provider credentials and other platform settings are configured in FleetQ, including AI provider keys, outbound connectors, and webhooks. The documentation describes encrypted credential storage, HMAC webhooks, tenant isolation, and SSRF protections. The provided material does not specify required environment variable names or a complete Docker Compose command, so deployment details should be taken from the repository's current setup instructions.

Tools and capabilities

The README describes hundreds of MCP tools across 45 domains. The available surface covers platform administration and agent operations, including:

• Creating and managing agents, skills, crews, experiments, projects, and workflows.
• Assigning tools and skills to agents and configuring model providers.
• Running workflow and crew executions with human approval steps.
• Inspecting task output, artifacts, timelines, metrics, errors, and audit records.
• Managing external MCP servers, built-in tools, integrations, signals, and webhooks.
• Reviewing execution history and proposing agent configuration changes through agent evolution.
• Observing agent teams through a live graph and OpenTelemetry-compatible tracing.

Limitations and notes

The escapeboy/agent-fleet-o MCP server is licensed under AGPL-3.0. The material confirms Claude Desktop and Cursor as clients that can drive the platform, but it does not provide client-specific configuration in the supplied excerpt. A hosted MCP endpoint is listed at https://fleetq.net/mcp; authentication requirements for that endpoint are not stated here.

The system is broader than a simple tool adapter: it includes a web application, orchestration runtime, storage, provider configuration, governance, and observability features. Using it therefore involves operating or accessing the FleetQ platform, rather than only installing an isolated MCP executable. Hosted model providers may require their own credentials, while the platform also supports local models.