agent-fleet-o

FleetQ is an open-source AI agent orchestration platform, officially positioned as a "self-hosted mission control center for AI agents," aimed at teams that need to run autonomous multi-agent systems in production environments, rather than toy projects meant only for demos. The repository is named escapeboy/agent-fleet-o.

[Purpose] FleetQ is used to build, run, and monitor autonomous multi-agent systems. It provides visual DAG workflow building, human-in-the-loop approvals, MCP server integration, and complete audit trails, and can be used with Claude, GPT-4o, Gemini, Ollama, Codex, Claude Code, and any OpenAI-compatible LLM.

[Main Capabilities]

1. MCP tools: Provides 670+ Model Context Protocol tools, covering 60+ domains, and any LLM (Claude Desktop, Cursor, ChatGPT, local agents) can programmatically drive the platform. Version 1.27 added a Web UI for capabilities that previously had no interface (agent sessions, publishing signing keys, drift and evaluation monitoring, broadcasts, test suites, CSV import), eight outbound chat channels as first-class drivers, as well as the Agentic AI Flywheel (self-growing evaluation sets + drift/production monitoring), policy-governed autonomy (versioned policies per agent + replay), cost-aware orchestration, and ROCS (Return on Cognitive Spend) metrics.
2. Visual DAG workflows: 8 node types (agent, conditional, human-task, switch, dynamic-fork, do-while, compensation, sub-workflow), with no Python glue code required.
3. Multi-agent teams: Supports coordinator/worker/reviewer roles, weighted QA scoring, and cross-validation.
4. Real-world action governance: Assistant tool calls, integration writes, and git push are routed through policies by risk level (low/medium/high corresponding to auto / ask / reject), and after approval are executed automatically with an audit trail.
5. Budget control: A real credit ledger, pessimistic locking, and automatic suspension on overspending.
6. Agent evolution: LLMs analyze execution history and propose configuration changes, with one-click approval.
7. BYOK and local LLM: Anthropic, OpenAI, Google, as well as Ollama, LM Studio, vLLM, Codex, Claude Code, avoiding vendor lock-in.
8. Production-grade features: tenant isolation, encrypted credential vault, HMAC webhook, SSRF protection, circuit breaker, audit trail.
9. OpenTelemetry observability: structured error codes (gRPC specification), deadline propagation, distributed tracing, Jaeger UI started with one command, per-team OTLP collection endpoint.
10. Real-time team graph: Cytoscape.js-based force-directed visualization, updated in real time via Laravel Reverb WebSockets.
11. Self-hosted or cloud: AGPLv3 licensed, can run with Docker Compose, or use FleetQ Cloud.

[Feature Sections]

• Agents, teams, and workflows: AI agents (role, goal, backstory, personality traits, skill assignment, per-agent provider/model fallback chain); 14 pre-built agent templates (engineering, content, business, design, research); agent evolution; agent teams (coordinator/QA/worker roles, 7 process types: sequential, parallel, hierarchical, self-claim, adversarial, fanout, chat-room, weighted QA scoring); pre-execution reconnaissance phase; step budget awareness; 20-state experiment pipeline; visual workflow DAG (including pre-built Web Dev Cycle template, NL-to-workflow generator); projects (one-off and continuous projects, cron scheduling, budget caps, milestones, overlap policy).
• LLM and compute: BYOK; local LLMs (Ollama, LM Studio, vLLM, llama.cpp, 17 preset Ollama models, SSRF protection); local agents (Codex, Claude Code, auto-detection, zero cost); Portkey Gateway (optional, unlocks 250+ LLM providers); RunPod GPU integration; pluggable compute providers (RunPod, Replicate, Fal.ai, Vast.ai); AI gateway (PrismPHP, 6-layer middleware, circuit breaker,fallback chain); semantic cache (pgvector, cosine similarity threshold 0.92).
• Signals, triggers, and outbound: 20+ signal connectors (webhook, RSS, IMAP, Slack, Discord, WhatsApp, GitHub, Linear, Jira, PagerDuty, Sentry, Datadog, Telegram, Matrix, Notion, Confluence, Screenpipe, Searxng, etc.); bug report signal (public JS component, screenshot + console + network + action log capture, threaded comments, agent delegation, SLA escalation); trigger rules (event-driven automation, condition evaluator, dry-run testing); multi-channel outbound (Email SMTP, Webhook, ntfy, and eight chat channels: Telegram, Slack, Discord, Microsoft Teams, Google Chat, Matrix, Signal, Supabase Realtime); Webhook (inbound HMAC-SHA256, outbound retries and event filtering).
• Human-in-the-loop, budget, and security: approvals (inbox + SLA enforcement and escalation); human tasks (embedded form mode within workflow nodes); credit ledger (per experiment and project, pessimistic locking, automatic pause on overspend); credential vault (encryption, rotation, OAuth2, expiration tracking, per-project injection); SSH tool (TOFU fingerprint verification, per-tool allowed command whitelist, multi-layer command security policy); audit trail; tenant isolation.
• Integrations and web development pipeline: integrate GitHub, Slack, Notion, Airtable, Linear, Stripe, Vercel, Netlify, etc.; autonomous web development pipeline (open PR, merge, dispatch CI, create release, trigger Vercel/Netlify/SSH deployment); website builder (AI-generated static sites, 8 component types, Vercel + ZIP deployment drivers); Founder Mode package (6 role agents, 20 framework skills, 5 pre-built workflows); marketplace (browse, publish, install shared skills/agents/workflows/packages, with AI risk scanning).
• API and MCP: REST API (175+ endpoints under /api/v1/,Sanctum authentication, cursor pagination, OpenAPI 3.1 documentation); MCP server (670+ tools, stdio + HTTP/SSE + OAuth2/PKCE); real-world action governance; public discovery endpoint /.well-known/fleetq; live team graph; structured MCP errors (gRPC-style error codes and retryable hints); per-tool deadlines; OpenTelemetry tracing; tool management; MCP client compatibility (Claude Desktop, Claude.ai, ChatGPT Apps, Cursor, Codex, Claude Code, Gemini CLI, etc.); queue management (Laravel Horizon, 6 priority queues); testing (regression test suite); per-invocation working directory.

[How to Use] Docker Quick Start: copy .env.example to .env; build and start all Docker services; run the interactive installation wizard (database, administrator account, LLM provider); after completion, visit http://localhost:8080。 Manual Web Setup: requires PHP 8.4+, PostgreSQL 17+, Redis 7+, Node.js 20+, Composer; open http://localhost:8000，设置页面引导创建管理员账户；也可运行 php artisan app:install interactive CLI wizard and seed default agents and skills. The self-hosted version skips email verification, and registration activates immediately; all registered users are automatically added to the default workspace. For local installation, you can set APP_AUTH_BYPASS=true to enable passwordless mode (official warning: never enable this on a server accessible from the internet). Configuration is centralized in .env; more LLM keys can be configured after logging in at Settings > AI Provider Keys; local models are configured at Settings > Local LLM Endpoints. Agents can execute commands on the host or remote servers through the built-in SSH tool type.

[Use Cases] Autonomous development pipelines, customer support triage, multi-agent research, scheduled content operations, incident response, GPU workloads, local-first agent development, and using FleetQ as an MCP server for Claude Desktop/ChatGPT/Cursor programming-driven use.

[Limitations] The official text does not provide specific performance metrics, pricing, concurrency limits, or similar data; such information is not yet available. The self-hosted license is AGPLv3. The warning about passwordless mode must be strictly observed.