Multi-agent AI that builds full-stack apps. Local-first via Ollama. Watch your agents dive, discuss, code, and ship — all visualized as a deep-sea expedition.
You type a plaintext prompt like "A markdown note-taking app with local search" and hit Launch Dive. Four AI agents (Archie — Planner, Byron — Backend, Faye — Frontend, Sentry — Security) run through a 9-round protocol:
- Analysis — Archie breaks the prompt into a technical brief
- Discussion — all agents align in parallel
- Consensus — directive locked in
- Architecture — Archie emits a blueprint (stack, routes, schema)
- Backend — Byron writes FastAPI + models + auth
- Frontend — Faye writes React + Tailwind + components wired to Byron's routes
- Security — Sentry audits, fixes high-severity issues in place
- Deployment — Dockerfile, docker-compose,
run.sh,README.md - Finalize — everything written to disk + Obsidian vault + knowledge graph
Output: a complete runnable project folder. Click Run App in the UI to launch it in Terminal. Click Open Brain in Obsidian to explore the reasoning vault.
- macOS (tested on Apple Silicon)
- Ollama running locally
- Node.js 18+, Python 3.11+, Rust toolchain, uv
ollama pull qwen2.5-coder:7b
ollama pull phi4-mini
ollama pull nomic-embed-text # for AI/ML analyticsnpm install
cd sidecar && uv sync && cd ..npm run tauri devTauri launches the app, which auto-spawns the FastAPI sidecar on :8765. First build primes the models (~5s warm-up hidden behind prompt typing).
echo "ELEVENLABS_API_KEY=sk_..." > sidecar/.envWithout a key, agents still emit dialogue as subtitles.
- Deep sea stage: animated reef — fish shoals, jellyfish, kelp, coral, god-rays, bubbles
- Agents: four squid-shaped minions arranged around a central orchestrator squid
- Live code terminals: click any agent while they're working → their token stream opens in a corner panel
- Build timeline: right panel shows phase progress
- AI/ML HUD: header badge shows embedding-based consensus % and code-quality score, computed live via
nomic-embed-text - Timing HUD: header shows elapsed build time and prompt→security latency
- Voice subtitles: agents speak during rounds (ElevenLabs TTS if configured)
┌──────────────────────────────────────────────────┐
│ Tauri desktop app (React + Tailwind + Motion) │
│ ────────────────────────────────────────────── │
│ EventSource stream ← SSE ← Python sidecar :8765 │
└──────────────────────────────────────────────────┘
↓
┌──────────────────────────────────────────────────┐
│ FastAPI sidecar (orchestrator.py) │
│ • Archie → qwen2.5-coder:7b (planning) │
│ • Byron → qwen2.5-coder:7b (backend) │
│ • Faye → qwen2.5-coder:7b (frontend) │
│ • Sentry → qwen2.5-coder:7b (security) │
│ • Orch → phi4-mini (dispatcher + voice) │
│ Embeddings: nomic-embed-text │
│ Voice: ElevenLabs turbo v2.5 (optional) │
└──────────────────────────────────────────────────┘
↓
Ollama runtime on :11434
For deeper detail see HOW_IT_WORKS.md.
YourApp/
├── run.sh ← one-shot launcher (chmod +x, runs installers + dev servers)
├── README.md ← quickstart + API reference + stack
├── ARCHITECTURE.md ← system design doc
├── Dockerfile ← production image
├── docker-compose.yml ← full stack compose
├── .gitignore
├── .env.example
├── requirements.txt ← backend deps
├── package.json ← frontend deps
├── main.py ← FastAPI app
├── routes/ ← API endpoints
├── models.py ← SQLAlchemy models
├── src/ ← React components
└── .brain/ ← Obsidian vault
├── 00-Index.md
├── Agents/ ← full transcript
├── Architecture/ ← decisions + rationale
├── API/ ← route catalog
├── Components/ ← frontend index
├── Database/ ← schema notes
├── Security/ ← Sentry's audit
├── Files/ ← per-file stubs (Obsidian-linked)
└── graph.html ← Graphify knowledge graph
Open .brain/ as an Obsidian vault (File → Open folder as vault) for full navigability.
If you Launch Dive into a folder that already has a .brain/ vault, OctoForge shows an "Existing project detected" banner with a one-click button to prepend this to your prompt:
"Before making changes, read the Obsidian vault at .brain/ to understand the existing architecture, API routes, and security audit. Then: "
Agents then read the vault first and avoid clobbering prior decisions.
Set in sidecar/agents/base.py:
"num_ctx": 4096 # halved vs default for ~2× throughput
"num_batch": 1024 # large batch for prompt eval
"num_thread": 8 # M-series perf cores
"top_k": 10 # tighter sampling
"keep_alive": -1 # keep models resident until sidecar exitsParallel agent round (Discussion) cuts a typical 4-step chain to 1 wall-clock step.
Frontend: Tauri 2, React 18, TypeScript, Tailwind CSS, Framer Motion, Zustand, Lucide icons Backend: Python 3.11, FastAPI, httpx, uvicorn Models: Ollama (qwen2.5-coder:7b, phi4-mini, nomic-embed-text), optional Claude Haiku via Anthropic SDK Voice: ElevenLabs Turbo v2.5 Knowledge graph: Graphify Vault: Obsidian-compatible markdown
MIT — hack away.
