AgentCafe is an open marketplace where AI agents discover real-world services, open a Tab, and place orders — with a human in the loop for anything that matters.
Your AI agent connects to AgentCafe through the Model Context Protocol endpoint. No SDK required — just point and go.
Your agent checks what's available — hotels, food delivery, home services, and whatever gets added next.
Want to book something? The human behind the agent reviews and approves it first — one tap with a passkey.
The Cafe validates every request, proxies it to the backend, and logs the whole thing. Two checks on every order.
Every action your agent takes on a real service needs your approval. You see exactly what it wants to do, for how long, and you can revoke access at any time.
Company Cards give your agent standing access to services you trust. Set budgets, exclude actions, and monitor usage from your dashboard.
Bring your OpenAPI spec and the onboarding wizard generates your Menu entry, configures auth, and sets risk tiers — live in minutes.
Any MCP-compatible AI agent can discover and use your service. No SDK integration, no per-agent onboarding. Just list it on the Menu.
Search and book hotel rooms worldwide. Find availability by city, dates, and preferences. Book rooms, view details, and manage cancellations.
Order food delivery from nearby restaurants. Browse menus, place orders, track deliveries, and manage cancellations. Delivers within 30 minutes in supported areas.
Find and book home service professionals — plumbing, electrical, cleaning, HVAC, handyman, painting, and landscaping. Search available providers, book appointments, reschedule, or cancel.
Human-owned persistent memory for AI agents. Agents store and retrieve JSON documents; humans see everything, control access via scoped tokens, and can delete anything.
AgentCafe is under active development. Some services on the Menu are live production services; others are demos with simulated backends. Check each service's description for details. APIs and data formats may change without warning during the beta period.