LeBonCoin has no good conversational search. You type keywords, scroll through noise, and miss deals. PolyAgent lets you chat naturally — "I'm looking for a 3-room apartment in Lyon under 800€" — and the agent refines, searches, and shows you exactly what matches.
Connecting an LLM agent to a real search backend via MCP stdio, handling streaming responses with tool calls, and making the conversational flow feel natural while the agent toggles between asking questions and fetching results.
FastAPI + SSE streaming
Server-Sent Events keep the conversation flowing in real time. The backend streams LLM tokens and tool results as they arrive — no polling, no WebSocket complexity.
MCP stdio for LeBonCoin
The agent talks to leboncoin-mcp via JSON-RPC over stdio — a clean separation between the agent logic and the search backend. Easy to swap for other platforms.
OpenAI-compatible API
Works with any OpenAI-compatible endpoint (9router, Gemini, Claude). Tool calling is handled at the API level — the agent just defines tools and the LLM decides when to call them.
Vanilla frontend + i18n
No framework overhead. HTML/CSS/JS with marked.js for Markdown rendering. Trilingual (FR/EN/ES) from day one with a simple locale object.
Working conversational agent with real-time search, clickable choices, favorite saves (localStorage), persistent sessions, and trilingual support. Deployed locally, extensible to any OpenAI-compatible API.
- MCP stdio is a clean abstraction for connecting LLMs to external tools — the protocol is simple and the separation of concerns is real.
- SSE streaming makes conversational AI feel instant — the user sees tokens appear as the model generates them, not after a full response.