LaRuche

Your personal AI hive. Local first, auditable, built in Rust.

LaRuche is a desktop and server application for running an AI agent on your own machine. It combines a resilient agent engine, cognitive memory, supervised automation, native computer and browser control, voice, messaging channels and a local mesh.

infinition.github.io/LaRuche

Quick start · Capabilities · Architecture · Wiki · Changelog

---

What is LaRuche

LaRuche is a self-hosted AI agent written primarily in Rust. The desktop application opens the web interface in its own window and starts the local node when needed. The same node can run alone on a server, and the lightweight desktop client can discover and use another hive on the local network.

The core works with llama.cpp, Ollama, LM Studio, vLLM, OpenAI-compatible endpoints, Anthropic and Codex authentication. Local models are the default use case, but cloud providers remain available when wanted.

The interface, memory, sessions, skills and configuration stay on your machine. The web assets are bundled, the node binds to loopback by default, and the default memory is a SQLite database that can be inspected and backed up with ordinary tools.

What LaRuche does

Agent engine

The butinage engine is a hardened ReAct loop built for the uneven tool-calling behavior of local models. It supports native provider calls with tolerant fallbacks, JSON Schema validation, token and tool-output budgets, per-tool timeouts, cooperative cancellation, live steering, anti-loop detection, context compaction and parallel scout agents.

An eval harness runs fixed missions against the assembled engine, not a simplified mock, so changes can be measured against saved baselines.

Cognitive memory

Memory is a graph of nodes and facts over SQLite, FTS5 and optional embeddings. Recall combines semantic and full-text results. Importance decays, updated facts supersede stale ones, and only memories used in an answer receive hebbian reinforcement.

The map can be exported to markdown in its own git repository. git log becomes a learning timeline, git diff shows what changed, and a previous snapshot can be re-imported without replacing the whole database.

Computer, browser and images

The native computer tool can inspect the Windows accessibility tree, capture any monitor, move or resize windows, and operate the mouse, keyboard and clipboard. A visible halo shows automated actions, an elevation check explains blocked input, and Ctrl+Alt+Shift+H is the emergency stop.

The dedicated Chrome extension connects LaRuche to the browser you already use, including its open tabs and signed-in sessions. The browser tool handles frames, shadow DOM, overlays, uploads, downloads, dialogs, touch emulation and responsive viewport checks. Consent banners are reported, never accepted automatically. Its data handling is documented in the Chrome extension privacy policy.

Images can be pasted, dropped or attached in chat. Tool screenshots and webcam captures reach vision-capable models, and oversized PNG or JPEG inputs are resized before they hit provider limits.

Supervision and automation