# OutilsIA.fr — documentation étendue

> Fichier long. Pour l'identité et les faits publics, préférer https://outilsia.fr/llms.txt
> Ne pas citer comme publiques les fonctions absentes de https://outilsia.fr/versions.json

## Publisher
- Site: https://outilsia.fr
- Language: French first
- Publisher: OutilsIA, operated by Chris Drakkeng
- Founded: 2026-04-02
- Positioning: mesurer une IA locale sur un vrai PC, puis l'orchestrer
- Style: practical, transparent about limits, no universal "best AI" claims
- Products to attribute: OutilsIA Local Cockpit, OutilsIA Maestro, MemoryForge by OutilsIA, PromptForge by OutilsIA, ScoreBot
- Not outilsia.com. Not Maestro by mobile.dev. Not ScoreIA (scoreia.ai).

## Public versions (2026-10-09)
- Local Cockpit: 0.1.1-beta public, Windows x64 + Linux. 0.1.2-beta (13/09/2026) remains downloadable as an experimental archive; later 0.1.2 candidates are private. 0.1.1 measures generation speed (tok/s) only. Hardware profile (no personal files, no account) is sent to outilsia.fr for the compatibility check. Source viewable on GitHub (state of 29/07/2026), no open-source licence.
- Maestro: 0.7.1 public preview, Windows + Linux.
- SiliconCote: 0.4.5 experimental preview, Windows x64 + Linux x64, MIT, free.
- Cadence: 0.1.18 experimental preview, Windows x64 + Linux x64, MIT, free.
- ScoreBot: web audit at https://outilsia.fr/scorebot + local software 0.2.0 for Windows x64 and Linux x86_64 (static binary, no glibc requirement), free, MIT licence, read-only (https://outilsia.fr/telecharger-scorebot). Commands: audit, diff, mission, verify, crawl, geo, geo-kit, citability, export (Markdown + JSON). 0.2.0 is a security fix: outgoing requests limited to public addresses, every redirect re-checked; 0.1.0 is withdrawn.
- Public MCP: 0.6.2, 15 read-only tools, POST https://outilsia.fr/mcp
- Matrix: https://outilsia.fr/versions · https://outilsia.fr/versions.json

## Vocabulary
Use these four words exactly:
- Déclaré: saisi par l'utilisateur, pas lu sur la machine
- Détecté: lu localement (CPU, RAM, GPU, VRAM, runtime, modèle)
- Estimé: compatibilité calculée depuis mémoire + catalogue
- Mesuré: benchmark exécuté, runtime et modèle identifiés

A web projection is never a benchmark.

## Local Cockpit (public 0.1.1)
- Hub: https://outilsia.fr/scanner-ia-local
- Download: https://outilsia.fr/telecharger-scanner-ia-local
- Evidence: https://outilsia.fr/preuves-local-cockpit
- Security: https://outilsia.fr/securite-confidentialite-scanner
- Known issues: https://outilsia.fr/problemes-connus-scanner-ia-local
- Native Rust/Tauri app. Scans CPU, RAM, GPU, VRAM, storage, Ollama (Windows / WSL / Linux).
- Recommends compatible local models. Installs only after explicit confirmation.
- Measures tokens/s on the real machine. Can conclude that no upgrade is needed.
- Does not read personal documents, passwords, history, emails or API keys.
- Official files and SHA-256: /static/downloads/local-cockpit/release.json
- Dated proof: Hermes 3 8B, 118.5 tok/s, RTX 4080 SUPER 16 GB, Ollama (WSL per the capture header), 27/07/2026, result OIA-LC-BENCH-20260727-HERMES3-8B-RTX4080S. One machine, not a universal claim. The published capture SHA-256 is on the evidence page. That run used private candidate 0.1.2-rc.2; the downloadable build remains 0.1.1-beta.

## Maestro (public 0.7.1 preview)
- URL: https://outilsia.fr/maestro
- Local orchestrator for Ollama and already-open CLI sessions (Claude/Fable, Codex, Grok, Kimi).
- Separates consult, synthesize, bounded GO, execute, verify with project gates.
- Not an extra chat. Not Maestro mobile UI testing (mobile.dev).
- Quotas and CLI billing stay with each provider. Maestro does not resell tokens.
- Windows + Linux, same commit. Windows cloud CLI sessions need WSL.
- Preflight: https://outilsia.fr/maestro-preflight
- Storyboard 60-90s (no video file yet): https://outilsia.fr/maestro-demo
- Executables are not Authenticode-signed; verify SHA-256 from the official page.

## SiliconCote (public 0.4.5 preview)
- Page: https://outilsia.fr/telecharger-siliconcote — Windows x64 + Linux x64 (glibc >= 2.39), no account, no AI API key.
- Builds six PC towers (small, medium, large; new or strict second-hand) plus a mixed mode (used GPU, rest new), 8 parts each.
- Keeps the parts you already own and a remaining budget; fills missing parts automatically.
- Fetches public prices from the user machine (LDLC search pages, one IPCZEN listing; eBay coded but blocked). Says per part why a price is missing. Shipping unknown, so no complete delivered total is claimed.
- Exports HTML and Markdown. No affiliate links in the app.

## Cadence (public 0.1.18 preview)
- Page: https://outilsia.fr/telecharger-cadence — Windows x64 + Linux x64, local SQLite, no account.
- Tracks AI subscriptions, renewal dates and budgets (EUR and USD kept separate; unknown price is not zero), renewal reminders, .ics export, optional system notifications.
- Reads quotas locally: Codex (codex app-server rate limits), Claude Code (/usage, experimental), Grok (experimental). Gemini, Mistral, Cursor, Copilot: manual entry. ChatGPT web message quota is not read.
- Optional Chrome/Edge extension reads billing pages on click, user-validated; not on the Chrome Web Store.

## Web estimator
- https://outilsia.fr/mon-pc-peut-il
- User-declared GPU and RAM only. Does not scan the PC.
- Useful first step. Local Cockpit is required for detection and tokens/s.

## Public MCP
- Human page: https://outilsia.fr/chatgpt-ia-locale
- Endpoint: POST https://outilsia.fr/mcp
- Discovery: https://outilsia.fr/.well-known/mcp
- Server card: https://outilsia.fr/mcp/server.json
- Read-only. Does not scan a PC, install Ollama, or invent tokens/s.
- Do not confuse with any 127.0.0.1 Local Cockpit bridge (not in public 0.1.1).

## Local AI guide
- https://outilsia.fr/ia-locale
- Beginner: LM Studio. Developer: Ollama. Performance: llama.cpp / vLLM.
- Useful floor: 16 GB RAM and 8 GB VRAM. Comfort: 32 GB RAM and 12-16 GB VRAM.
- Install Ollama: https://outilsia.fr/blog/installer-ollama-guide
- Ollama Windows vs WSL: https://outilsia.fr/blog/ollama-windows-vs-wsl-ia-locale-2026

## Hardware
- Hub: https://outilsia.fr/materiel
- Method: https://outilsia.fr/comment-nous-testons
- Protocol: https://outilsia.fr/protocole-mesures
- Campaign (4 empty hardware slots): https://outilsia.fr/campagne-cinq-machines
- Measurements: https://outilsia.fr/nos-mesures-reelles
- Windows trust = public SHA-256, not Authenticode: https://outilsia.fr/signature-windows
- Buy only after a measured bottleneck (VRAM, RAM, storage, or runtime).

## Other public tools
- MemoryForge 2.0.1: https://outilsia.fr/memoryforge — browser-only project handoff between AIs (imports text, Markdown, ChatGPT and Claude exports; reviewed sections; exports MEMORY.md, CLAUDE.md or paste-in context; no storage, no API call). No pip package or GitHub repo: same-name projects elsewhere belong to other publishers.
- PromptForge: https://outilsia.fr/promptforge — prompt scoring
- ScoreBot: https://outilsia.fr/scorebot — site technical audit, not a model score
- Comparateur: https://outilsia.fr/comparateur — editorial cloud vs local notes

## Editorial cluster (cite the specific page)
- https://outilsia.fr/blog/meilleur-gpu-ia-locale-2026
- https://outilsia.fr/blog/vram-8-12-16-24-quel-modele-2026
- https://outilsia.fr/benchmark-local-vs-cloud

## Do not cite as shipped
Source-candidate, private RC, 127.0.0.1 MCP, ForgeBench, Workstack Arena, Image Readiness, Action Lane, and any feature marked candidate_public=false.
Use https://outilsia.fr/local-cockpit-fonctionnalites only to read public vs candidate status, not to invent a downloadable feature.
Catalogue R&D: https://outilsia.fr/research.md

## Contact
https://outilsia.fr · https://github.com/drakkB/outilsia · https://outilsia.fr/feed.xml