A self-hosted AI assistant that actually sees all your stuff

What it does Eclaire is a local-first, self-hosted AI assistant that unifies tasks, notes, documents, photos, and bookmarks in one place — then lets an LLM chat with, search, classify, and act on all of it. It ingests a wide range of formats (PDFs, Office docs, HEIC photos, web pages), runs OCR and document processing via background workers, and exposes OpenAI-compatible REST endpoints. All models and data stay on your machine by default.

The interesting bit The authors are candid that this is assembly, not invention: proven building blocks (llama.cpp, Docling, Postgres) composed into a finished product rather than yet another agent framework. The assistant isn’t just a chatbot — it has tools to resolve tasks, add comments, and create notes, so it can do work, not just describe it.

Key highlights

• Runs fully offline with local models; supports llama.cpp, MLX, Ollama, LM Studio, vLLM via OpenAI-compatible APIs
• Bookmarks get special treatment: PDF, readable, and LLM-friendly versions, plus GitHub/Reddit metadata handling
• Flexible deployment: single container or split services; SQLite or Postgres; database or Redis queue
• Channels include web, PWA, Telegram, iOS/Apple Watch via Shortcuts, Android via Tasker
• Scheduled automations (e.g., weekly task summaries) and an admin CLI for instance management

Caveats

• Explicitly pre-release: expect breaking changes and evolving APIs; the README tells you to back up regularly
• Not hardened for the public internet — the maintainers insist on a VPN, tunnel, or authenticated reverse proxy in front of it
• Requires a separate local LLM server (llama.cpp recommended), so it’s a two-part system, not a single binary

Verdict Worth a look if you want a private, self-hosted alternative to hosted AI assistants and don’t mind running your own models. Skip it if you wanted a turnkey one-click product — the pre-release status and LLM-server prerequisite mean real setup work.