agentpack
Define your agent team in YAML — get a supervisor, a live network UI, an MCP server, and behavioral evals out of the box.
agentpack is a batteries-included framework for multi-agent AI teams: a supervisor delegates to specialist agents, each owning a subset of tools, with everything around them — streaming UI, REST API, MCP endpoint, eval harness — generated for you.
Quick start
npx @selvaonline/agentpack init my-team
cd my-team && npm install
cp .env.example .env # add ONE key: OpenAI, Gemini, or Groq
npm run dev # → http://localhost:3000
Or let AI design the whole project from a description — manifest, prompts, tool stubs with sample data, and evals:
npx @selvaonline/agentpack init --describe "a restaurant site-selection team: \
analyze foot traffic and demographics, score competition, recommend the best site"
You now have a working team with a live network view at http://localhost:3000. Prove it behaves:
npm run eval
What you get from one manifest
| Capability | How |
|---|---|
| Supervisor + specialist orchestration | LangGraph under the hood, configured from your YAML |
| Live network UI | hop-by-hop animation, token streaming, light/dark themes |
| Build agents in the browser | compose teams from the tool catalog — or describe one and let AI design it |
| MCP server | every team auto-exposed at /mcp for Claude, Cursor, any MCP client |
| Behavioral evals | routing, completeness, refusal, tool usage + LLM-as-judge |
| Guardrails | anti-rationalization excuse → rebuttal tables compiled into every prompt |
| Human-in-the-loop | approval: true pauses delegations for your sign-off |
| Embeddable widget | one script tag drops the live network panel into any page |
Where next
- The manifest — every field of
agentpack.yaml - Templates — seven ready-made vertical teams
- AI-assisted building — describe it, get a team
- HTTP API — run teams programmatically