Follows the rules
Eligibility and refund limits come from tested code, never from the AI's guess.
Agent · Tools · Memory · Human approval
The returns agent that knows when to ask.
An AI support agent for a demo store. It looks up orders, checks the return policy and starts returns. Before any risky refund, it stops and waits for a person to approve.
3-minute demo · no sign-up · each visitor gets a private copy of the store
Chat · Maya Patel
Waits for a human
Refunds over $50 pause until a reviewer approves, edits, or rejects them.
Remembers you
Preferences carry across chats, and every step is traced so you can see why.
How it works
One request, five safe steps
Every chat turn runs through the same pipeline. The interesting part is step four: the agent can't move money on its own.
1. Ask
A customer asks about an order in plain words.
2. Tools via MCP
The agent looks up orders and checks eligibility with real tools.
3. Policy engine
Tested rules — not the AI — decide what's allowed.
4. Human approval
Risky refunds pause until a reviewer approves.
5. Async job
An idempotent job issues the refund; chat updates live.
Architecture
Small, real, and free to run
A Next.js front end on Vercel talks to one Docker container on Render: the agent, the MCP server, the policy engine and a job queue, all backed by Postgres on Neon.
Browser
Chat · reviews · traces
Next.js on Vercel
Auth.js · route guards
JWT proxySSE passthrough
FastAPI on Render · one Docker container
LangGraph agent
Checkpoints · interrupt()
MCP server
Order & refund tools
Policy engine
Deterministic rules
Celery + Redis
Idempotent jobs
Groq + Gemini
Free LLMs · routing, fallback, quota guard
Neon Postgres
+ pgvector
Business dataMemoryTraces
What it shows
What companies actually build
Not a single prompt with no tools — the pieces a production support agent needs.
Real tools over MCP
A custom Model Context Protocol server exposes the order system. The customer id is injected from the token, so the model can't ask for someone else's data.
LLM proposes, code disposes
Write tools return proposals. A deterministic policy engine with 100% test coverage decides eligibility, limits and when a human is needed.
Pause, approve, resume
LangGraph interrupts at the approval gate and checkpoints to Postgres — a reviewer can decide hours later and the run resumes exactly there.
Memory you control
Long-term preferences live in pgvector, are filtered for sensitive data, and can be viewed and deleted by the customer.
Every step traced
Router, LLM calls, tools, guards, policy, interrupts and jobs — with timings and tokens — on a step-by-step timeline.
Evaluated on trajectories
30 scenarios check the right tools in the right order, forbidden tools never called and zero policy violations.
- Next.js 16
- React 19
- TypeScript
- Tailwind CSS v4
- Auth.js
- TanStack Query
- FastAPI
- LangGraph
- MCP
- Celery + Redis
- PostgreSQL
- pgvector
- Groq · Llama 3.3
- Gemini
- Vercel
- Render
- Neon
Play both sides of a risky refund.
Ask for a refund as Maya, approve it as Riley, then watch the job finish and the chat update — in about three minutes.