🧭 Компас — a negotiation travel agent for trips across Russia

An agent that a family from Yekaterinburg writes to in a single phrase:

"We want to go to the sea, 45,000 for three, with a child, without long connections."

— and it itself finds the contradiction between budget and connections, shows the fork right on the timeline, asks what to sacrifice, remembers the answers, and keeps Plan B ready for every fragile leg of the route.

Project for the Tutu MCP AI hackathon, August 18–21, 2026. Works on top of Tutu MCP (planes, trains, commuter trains, buses, hotels).

From left to right: the start exploration map (clicks on cities don't spend tokens — only Tutu MCP); the city card with flights and ±day; switching to Plan B after a failure; MCP widget (SEP-1865).

How this differs from "yet another ticket search"

Quick Start

git clone https://github.com/svyatrunov/tutu-compass.git && cd tutu-compass

npm install

cp deploy/.env.example apps/server/.env

Fill in OPENROUTER_API_KEY in apps/server/.env and run:

npm run dev

Frontend — http://localhost:5173, API — http://localhost:8787 (Vite proxies /api).

Demo without internet and without keys — on recorded fixtures:

DEMO_MODE=replay npm run dev

Production entirely in Docker:

cd deploy && cp .env.example .env && docker compose up -d --build

Компас as an MCP server

The same agent is available as the MCP tool plan_trip — one call instead of manual orchestration of the 16 raw Tutu tools. Endpoint: http://localhost:8787/mcp after npm run dev (Streamable HTTP; on your own server — https://<домен>/mcp). To quickly view the widget without an MCP host: http://localhost:8787/api/dev/widget-preview. In hosts with MCP Apps support (Claude, Claude Desktop, VS Code Copilot, Goose), the result is rendered as an interactive timeline right in the chat (SEP-1865); other MCP clients receive text + structured JSON. Dialog memory — via sessionId in the tool arguments. The web version is the primary channel and does not depend on the MCP channel: both live in one process on top of the same agent/* core.

Channel check: npm run smoke:mcpapp against a running server.

Useful Commands

Documentation

Where to look at the code by hackathon criteria

Repository Rules

• Secrets only via env. .env is in .gitignore, .env.example — in the repository.

• The LLM never generates route data: prices, flights, and links come only from MCP, the model only selects among them by id. This is protection against hallucinations on top of real data.

License

MIT — see LICENSE.