Build AI agent teams, not chatbots.
Self-hosted. Multi-agent. Observable.
AgentHub lets you spin up a team of AI agents that actually work together — Router figures out what needs doing, Executor does it, Critic checks the work, Summarizer ties it all up. Not a single-agent-with-tools trick. A real team.
Each agent runs through an 11-state ReAct loop that you can watch in real time. Everything streams. Everything logs. Everything's on your hardware.
Most AI platforms wrap an LLM in a chat box and call it an agent. Then you spend weeks wiring together "multi-agent" flows that break the moment something unexpected happens.
AgentHub gives you the full loop out of the box — orchestration, IAM, sandbox execution, search, observability. You bring a model key. It brings the rest.
git clone https://github.com/EVEDensity/AgentHub.git
cd AgentHub
start.batThat's it. PostgreSQL spins up via Docker, the backend starts, the frontend opens. No API key needed — the built-in mock provider lets you kick the tires offline.
Want your own model?
export OPENAI_API_KEY=sk-...
# or
export ANTHROPIC_API_KEY=sk-ant-...
# or any OpenAI-compatible endpoint
export OPENAI_COMPATIBLE_BASE_URL=http://localhost:11434/v1Then:
docker compose -f deploy/docker-compose.platform.yml up --build| Service | URL |
|---|---|
| Web UI | http://localhost:3000 |
| API Gateway | http://localhost:8081 |
| Grafana | http://localhost:3001 |
Agent orchestration — 6 roles (Router, Planner, Executor, Critic, Summarizer, Search) running a budget-aware ReAct loop with Redis-backed state that survives restarts. Streaming by default via WebSocket + SSE.
Security — JWT auth, RBAC + ABAC, sensitive-tool gating, per-tenant data isolation. Built for production, not just demos.
Search — BM25 + dense vectors + reranking in one pipeline. Graceful fallback when vector stores are down. Pluggable embeddings.
Sandbox — Code execution in isolated containers. Configurable CPU/memory limits, network policies, output sanitization.
Observability — Prometheus + Grafana + OTLP tracing. See what every agent is doing in real time.
Go services over NATS JetStream. Rust for performance-critical paths (stream processing, retrieval, fanout). Python for offline/async tasks (model adaptation, document pipelines, evaluation). Next.js frontend. PostgreSQL, Redis, Qdrant, MinIO underneath.
PRs welcome. Check good first issues for a place to start.
# Dev setup — infra only
cd deploy
docker compose -f docker-compose.platform.yml up -d nats postgres redis
# Go services
cd services/go && go work sync
# Frontend
cd frontend && npm install && npm run devBug reports, docs, new model providers — all count. CONTRIBUTING.md has the details.
Thanks to everyone who's contributed to AgentHub — code, docs, bug reports, ideas. It all matters.
Apache 2.0.
Built by Density and contributors.