LocalMate.

Agentic AI assistant for internal knowledge retrieval, built on MCP (Model Context Protocol). Targeted at any organisation that wants an overview of its own data.

Role: Co-founder, technical lead
Status: Live
Year: 2026
Links: localmate.ch
Live prototype

LocalMate is an agentic AI assistant that uses MCP (Model Context Protocol) as its central abstraction for knowledge and tool integration. The assistant is built for organisations that want to make their internal documents and processes accessible to staff and external stakeholders without handing sensitive data to third parties.

Architecturally, LocalMate combines a LangGraph.js ReAct loop with dynamically loaded MCP tools. The backend runs on NestJS atop Bun, the frontend is an Angular application. Retrieval runs against a Qdrant vector database; structured data lives in PostgreSQL, accessed through Drizzle ORM. The full codebase sits in a Turborepo monorepo with end-to-end TypeScript typing.

LocalMate won the Hack Winterthur 2026 hackathon. The live prototype at winti.localmate.ch currently handles roughly 600 queries per week. The product is privacy-first, single-tenant, and anonymous, with no user accounts.

We are currently looking for our first official partner organisations in Switzerland, across both the public and private sector, to deploy LocalMate in production.