Drive an installed LaRuche App through its declared actions, never its files.
日本語の概要は準備中です。原文の説明を表示しています。
What LaRuche is: architecture, butinage, memory, LaReine, and its full wiki.
インストールする前に、エージェントに与えられる指示の中身を確認できます。
Ce skill embarque le wiki complet de LaRuche, 34 pages, 184 Ko de markdown. C'est
la MEME source que le site publie: wiki/ a la racine du depot, dont docs/wiki.html
et ce dossier sont deux sorties generees par scripts/build_wiki.py. Ce qui est ecrit
ici fait donc foi. Ne repondez jamais de memoire sur le fonctionnement de LaRuche:
ouvrez la page et citez-la.
Toute question sur LaRuche elle-meme: ce qu'elle sait faire, comment elle est faite, ce qu'un mot du vocabulaire designe, pourquoi elle se comporte d'une certaine facon, ce qu'elle fait de vos donnees. Y compris la demande d'accueil "Presente-toi".
Ce skill dit ce que LaRuche EST. Pour FAIRE quelque chose, les skills voisins sont plus
directs: configure-laruche (reglages, fournisseur, canaux, secrets),
cognitive-memory (retenir et retrouver), extend-toolset (Forged Tool, MCP),
delegation (sous-agent), long-running-work (mission, kanban, plan).
LaRuche est un agent local: le butinage est sa boucle de raisonnement et d'action, les abeilles sont ses outils, les eclaireuses ses sous-agents, LaReine sa supervision, et sa memoire est une carte de faits au format OKF, lisible et modifiable a la main.
Ouvrez la page qui repond, avec file_read, sur le chemin donne relatif au dossier de
ce skill (skills/laruche/). Les corps sont integraux: ils ne coutent rien tant que
vous ne les demandez pas.
| Page | Titre | Ce qu'elle couvre |
|---|---|---|
| Overview | ||
wiki/Home.md | LaRuche Wiki | Welcome to the hive. LaRuche is a local-first AI agent with a desktop application, a server node and a lightweight LAN client. |
wiki/Tour.md | A tour of LaRuche | The rest of this wiki explains how LaRuche works. |
wiki/FAQ.md | FAQ | No. With llama.cpp or Ollama, a local embedding model and local speech backends, the engine, memory, automation and voice run... |
wiki/Security.md | Security | LaRuche's security model starts from an honest premise: the model will eventually do something wrong. |
| Getting started | ||
wiki/getting-started/Installation.md | Installation | The normal entry point is the desktop application. |
wiki/getting-started/Desktop-App.md | Desktop App | LaRuche ships as a desktop application, a server node and a terminal client. |
wiki/getting-started/Quick-Start.md | Quick Start | You have opened the desktop application or started the node (Installation). |
wiki/getting-started/Local-Models.md | Local Models | LaRuche is designed local-first, and specifically for the reality of local models: they are slower, they misformat tool calls,... |
| Concepts | ||
wiki/concepts/Architecture.md | Architecture | LaRuche separates the desktop shell, the server node and the terminal client. |
wiki/concepts/Butinage-Engine.md | The Butinage Engine | Butinage is LaRuche's agent loop. It turns a request into a sequence of model calls, tool calls, observations and decisions,... |
wiki/concepts/Cognitive-Memory.md | Cognitive Memory | LaRuche's memory is not a vector store bolted onto a chat log. |
wiki/concepts/LaReine.md | LaReine | LaReine is LaRuche's supervision layer. It can review a finished answer, send the worker through a fresh agentic run, gate... |
wiki/concepts/Table-Ronde.md | Table Ronde | The table ronde is LaRuche's structured multi-agent deliberation mode. |
wiki/concepts/Watchers.md | Watchers | A watcher is a standing condition. It observes something, decides whether that means anything, and reacts. |
wiki/concepts/Automation.md | Automation | The automation hub gathers everything the hive does on its own: crons, missions, the kanban, and watchers. |
wiki/concepts/Skills-and-Curator.md | Skills and the Curator | A skill is a markdown file: instructions, examples, and conventions for a category of task. |
wiki/concepts/Apps.md | Apps | An App is an installable package that brings its own interface into LaRuche, and can declare actions an agent is allowed to call. |
wiki/concepts/Forged-Tools.md | Forged Tools | A Forged Tool is an atomic command the agent can call like a built-in tool. |
| Guides | ||
wiki/guides/Computer-and-Browser.md | Computer and Browser | LaRuche can act on the desktop, the user's Chrome session and visual inputs. |
wiki/guides/Chrome-Extension.md | Chrome Extension | The LaRuche extension lets browser control the Chrome instance the user already has open, including its tabs and signed-in... |
wiki/guides/Training-Datasets.md | Training Datasets | LaRuche can turn completed LaReine reviews into training data. |
wiki/guides/Voice.md | Voice | LaRuche can speak and listen through local or remote speech backends. |
wiki/guides/Telegram.md | Telegram | Telegram is LaRuche's most mature remote channel: full agent runs from your phone, persistent per-channel memory, voice... |
wiki/guides/MCP.md | MCP | LaRuche sits on both sides of the Model Context Protocol: it consumes external MCP servers as extra tools, and it exposes... |
wiki/guides/Secrets.md | Secrets | The secrets vault exists for one reason: your API keys, tokens, and passwords should never enter a model's context. |
wiki/guides/Troubleshooting.md | When something does not work | Symptoms first. Each one below has been seen, and the cause is rarely where the message points. |
wiki/guides/Developing-Apps.md | Developing Apps | Apps can be developed and installed without rebuilding LaRuche. |
| Reference | ||
wiki/reference/Configuration.md | Configuration | Two layers: supported environment variables set at launch, and the Settings UI for everything that can change live. |
wiki/reference/Providers-and-Profiles.md | Providers and profiles | Most confusion about "which model am I actually talking to" comes from one idea that is never stated: **the active model is a... |
wiki/reference/Tools.md | Tools | A default node build registers over a hundred built-in tools. |
wiki/reference/App-Manifest.md | The App manifest | needs to know before running any of its code. |
wiki/reference/App-SDK.md | The App SDK | Every capability an App has reaches it through this one object. |
wiki/reference/Evals.md | Knowing whether an engine change helped | Prompts and loops are the easiest part of an agent to change and the hardest part to judge. |
wiki/reference/Brand-Glossary.md | Brand Glossary | The hive speaks French. The brand vocabulary is part of LaRuche's identity and stays French in the code, the UI, and the docs;... |
Citez la page d'ou vient votre reponse, l'utilisateur peut la relire. Une page absente de ce tableau n'existe pas: dites-le plutot que de l'inventer.
まだレビューはありません。使ってみた感想をお寄せください。
概要と使いどころ
Drive an installed LaRuche App through its declared actions, never its files.
日本語の概要は準備中です。原文の説明を表示しています。
Find academic papers on arXiv, with citation counts and BibTeX.
日本語の概要は準備中です。原文の説明を表示しています。
Render text or an image as ASCII art for terminal-friendly output.
日本語の概要は準備中です。原文の説明を表示しています。
Track RSS/Atom feeds and blogs via blogwatcher-cli.
日本語の概要は準備中です。原文の説明を表示しています。
Drive a real web browser: navigate, read, find, click, fill, screenshot
日本語の概要は準備中です。原文の説明を表示しています。
Measure a codebase: lines of code, language mix, and symbol lookups.
日本語の概要は準備中です。原文の説明を表示しています。