🍄 Luigi Codes

Premium local AI coding agent by Luigi Solutions. Private, powerful, yours.

Site: luigi-codes.vercel.app · Source: github.com/LuigiSolutions/luigi-codes

Luigi Codes is a VS Code extension that puts a full coding agent — chat, editor commands, an autonomous tool-running agent loop, codebase intelligence, long-term memory, and a self-improvement system — on top of models that run entirely on your machine. No API keys, no per-token bills, no code leaving your computer.

Why

• Ownership. Your assistant, your models, your data, your brand. No vendor can change the price, the model, or the terms underneath you.
• Privacy. Prompts, code, memory, and improvement data never leave localhost.
• Compounding. Every run is recorded, analyzed, and folded back into better planning prompts — and accumulates into a fine-tuning dataset that is yours.

Features

Requirements

1. VS Code 1.85+
2. A local inference server — Ollama (default) or LM Studio
3. At least one model pulled:

# recommended starting set
ollama pull qwen2.5-coder:7b       # strong small coder
ollama pull deepseek-coder:6.7b    # fast code generation
ollama pull nomic-embed-text       # embeddings for index + memory

Honest sizing note: small local models (7–13B) are genuinely useful but are not frontier-model equivalents. Quality scales with what your hardware can serve — qwen2.5-coder:32b or larger on a capable machine narrows the gap substantially. The router is built so you can swap upward without touching anything else. Optional: run chroma run (ChromaDB) on port 8000 for vector memory; Luigi falls back to local JSON automatically.

Install & Run (development)

cd luigi-codes
npm install
npm run compile
code .
# press F5 → "Run Luigi Codes" launches an Extension Development Host

The web app (desktop & mobile)

Run Luigi: Open Web App (Desktop & Mobile) from the command palette — Luigi starts a zero-dependency local HTTP server and gives you a URL that opens the full chat in any browser. To use it from your phone: enable luigi.web.allowLan, re-run the command, and open the phone URL it offers (same Wi-Fi). Every URL carries a per-session access token — without it the server shows a locked page.

No VS Code around? The web chat also runs standalone:

npm run web -- --endpoint http://localhost:11434 --provider ollama --lan

The browser talks only to this server; this server talks only to your local inference endpoint. Nothing leaves your machine (or, with --lan, your network).

Testing & audits

All of these pass with no model server running — that is the contract.

npm test                 # integration suite (T1–T25) in a real downloaded VS Code:
                         # activation, commands, webview CSP/nonce, tool registry,
                         # router fallback, plan parsing, brand-token emission
npm run audit:imports    # every import resolves; zero circular dependencies
npm run audit:brand      # every hex/rgba in src/, media/, package.json is on-palette
npm run audit:copy       # zero em dashes in user-facing copy (src/, api/, site/, media/)

The first npm test downloads a VS Code build into .vscode-test/ (~260 MB, cached).

Open the Luigi icon in the activity bar, or run Luigi: Open Chat from the command palette.

Commands

Configuration

Architecture

src/
├── extension.ts              activation, 11 commands, status bar, terminal chat
├── ui/designTokens.ts        Luigi Solutions brand tokens (single source of truth)
├── chat/chatPanel.ts         chat webview + activity-bar sidebar
├── chat/markdown.ts          the one markdown renderer (panel, web app, tests)
├── inference/modelRouter.ts  model registry · detection · routing · streaming
├── inference/streamText.ts   NDJSON/SSE stream parsing + stop-marker guard (vscode-free)
├── github/githubClient.ts    zero-dep GitHub REST client (vscode-free, both surfaces)
├── github/githubTools.ts     the agent's GitHub tools (writes behind approval)
├── web/webServer.ts          zero-dep HTTP server: responsive chat for desktop & mobile
├── web/standalone.ts         `npm run web` — the web chat without VS Code
├── agent/
│   ├── agentLoop.ts          the 5-phase agent (context→plan→approve→execute→verify)
│   └── tools/toolRegistry.ts the agent's 11 workspace-guarded tools
├── context/codebaseIndex.ts  symbols, imports, patterns, semantic retrieval
├── memory/memorySystem.ts    ChromaDB + local JSON task memory
├── improvement/selfImprove.ts failure taxonomy → prompt rules → fine-tune readiness
└── test/                     @vscode/test-electron + mocha integration suite (T1–T25)
scripts/
├── audit-imports.mjs         import resolution + circular-dependency audit
├── audit-brand.mjs           palette compliance audit (src/, media/, package.json)
└── audit-copy.mjs            em-dash-free user copy audit (src/, api/, site/, media/)

Data flow, one run: chatPanel → agentLoop.execute() → codebaseIndex + memorySystem (context) → modelRouter (plan) → approval modal → toolRegistry (execute, self-correct) → modelRouter (verify) → memorySystem.storeTask() + selfImprove.analyzeTask().

Brand

Every color in this extension is the Luigi Solutions palette, taken exactly from luigi-os/packages/ui/src/theme.css: canvas #0b0a09, surface #16140f, ink #f3efe7, gold #c9a86a, danger #e8796e, success #a3c585, warning #d9924a, info #8fb5c9 — with the signature gold-at-32% hairline, eyebrow tracking 0.15em, wordmark tracking 0.3em, and the house easing cubic-bezier(0.16, 1, 0.3, 1). Derived steps (elevated surfaces, gold light/dark) live only in src/ui/designTokens.ts.

Roadmap

• Inline completions (ghost text) provider
• Diff-preview apply for agent file edits
• ✅ Fine-tune export (JSONL) + one-command local LoRA recipe (Luigi: Export Training Data + TRAINING.md)
• Multi-model ensemble verification for high-stakes edits

© Luigi Solutions. Built to be owned.