Restaurant-booking agent demo: FastAPI + Claude tool-use loop + real MCP server + eval-driven-dev harness
A self-contained mockup for the "AI Agents for Enterprise" lightning session: a customer-facing chatbot that searches restaurants and books tables through real MCP tools, backed by a mock "Elasticsearch" (SQLite) restaurant index.
The read/write tool split is deliberate, not cosmetic: read tools are safe for the agent to call as often as it wants while reasoning; write tools are where the system prompt requires the agent to restate details and get explicit user confirmation before calling. That's the hook back into the full course's "making agents reliable in production" theme — this repo is the free lightning-lesson teaser for it.
"Harness" is the code around the model call — everything that turns a single LLM request into an agent that can act. Claude by itself just maps one input to one output; it doesn't call tools, check results, or decide to try again on its own. The harness is what closes that loop. Strip the harness away and you have a plain chatbot: one message in, one reply out, no ability to actually go look anything up or do anything. Add the harness and the same model can search a database, notice it needs more information, and act on it — that's the entire difference between "chatbot" and "agent" in this repo (see the "chatbot vs. agent" discussion earlier in this project's history: the UI is a chatbot, chatservice/main.py is what makes what's behind it an agent).
Concretely, the harness here is the while True loop in chatservice/main.py's /chat handler:
From the project README.
Add the radar badge to your README — it shows your project was picked up by MCP Radar and links to this page:
[](https://mcp.liqiwa.com/s/skandavivek--openagent.html)
Turn scattered notes, docs and transcripts into a queryable Markdown wiki — an LLM knowledge-base compiler with MCP access, no embeddings, self-hosted.
alxgntv/substack-api-mcpMCP server for Substack posts (FastMCP + substack-api-client)
sm18lr88/STT-MCPFast, local speech-to-text MCP server
simonw/mcp-explorerCLI tool for exploring an MCP server
sandeepbazar/ocm-mcp-serverAn MCP server that lets AI agents operate a multi-cluster Kubernetes fleet through an Open Cluster Management hub, with policy, approval, and audit between the model and your clusters.
Vladimir-Human/ru-marketplace-mcpДевять российских маркетплейсов и китайский Taobao как MCP-серверы: Wildberries, Ozon, Яндекс Маркет, Детский мир, Авито, Мегамаркет, Lamoda, DNS, Ситилинк. Плюс сравнение цен по всем сразу. Только чт
The top new MCP servers of the week, every Monday. No spam, unsubscribe anytime.