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 — the Orchestrator decomposes your goal and dispatches roles, code-review and verification agents check the work, and deploy/implementation agents finish it. Not a single-agent-with-tools trick. A real team.
Each agent runs under a bounded, observable 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.
Run this in PowerShell to install the latest checksummed GitHub Release for the current user. It does not require administrator access, npm, or Python:
irm https://github.com/EVEDensity/AgentHub/releases/latest/download/install.ps1 | iexOpen a new terminal, then verify the installation:
agenthub --helpOnly Windows x64 CLI binaries are published today. Other platforms are not yet claimed as supported.
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/v1Optionally route all providers through a unified gateway (new-api / one-api style) for multi-key failover, quotas and usage analytics:
export AGENTHUB_LLM_GATEWAY=newapi
export AGENTHUB_NEWAPI_BASE_URL=http://127.0.0.1:3000/v1
export AGENTHUB_NEWAPI_API_KEY=sk-agenthub-xxxxSee deploy/newapi/README.md for deployment, migration and verification (ADR-0104).
Then:
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 — 7 domain roles (Orchestrator, Architect, CodeGen, Review, Test, Implement, Deploy) running a bounded tool-call loop. State is persisted in PostgreSQL; streaming is delivered via WebSocket + SSE. Role roster is defined in app/services/agent_service.py (DEFAULT_AGENTS).
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.
Public benchmark score — the developer CLI execution loop has a citable, replayable pass rate on a Terminal-Bench-style acceptance-command suite: 8/8 (100%) with deepseek-v4-flash (2026-09-01). See benchmarks/public-scores.md for methodology and honest scope.
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.