A new open-source tool called Universal Context Pipeline (UCP-Local) lets you ground AI assistants in your own files without sending anything to the cloud. Released on June 16, 2026 under an Apache-2.0 license, it runs as a local Model Context Protocol server that indexes your folders and exposes them to Claude Desktop, Cursor, and LM Studio for fully offline retrieval.

Try It: Offline RAG in Your AI Assistant

If you already use Claude Desktop or Cursor, you can wire UCP-Local into your workflow in a few minutes. Install it with cargo install ucp-local, point it at a folder with ucp-local index ~/Documents/notes, then start the server using ucp-local serve. Add the binary to your Model Context Protocol config and the assistant gains a search_local_context tool that pulls answers straight from your notes, code, and PDFs, all on your own machine.

Why It Matters for Creators

Most retrieval setups send your documents to a hosted embedding API. UCP-Local keeps everything local, using Ollama for embeddings so your private project files never leave your computer. For writers, developers, and studios working under NDAs or on unreleased material, that privacy difference is the whole point. It also pairs naturally with the growing set of local-first creator tools, joining options like our coverage of local AI agents for Cursor.

Key Details

License: Apache-2.0, free and open source.

Clients: Claude Desktop, Cursor, and LM Studio (v0.3.17 and later) via standard MCP config.

Retrieval: Hybrid BM25 (SQLite FTS5) combined with vector search, plus tree-sitter code chunking and PII masking on by default.

File types: Markdown, plain text, PDF, and source files for Rust, Python, TypeScript, JavaScript, and Go.

Live updates: Watches directories and refreshes the index within roughly 500ms of a file change.

What to Do Next

UCP-Local is at version 0.1.0, so expect rough edges: it supports Ollama embeddings only, indexes Claude conversation exports but not Cursor or ChatGPT formats yet, and ships without a desktop UI. If you want private, offline grounding for an AI coding assistant right now, it is worth a test run on a single project folder before you point it at an entire workspace.