Recursive Evolution Protocol Loop Optimizing Intelligent Dreamer (Dynamic Recursive Engine Adapting Modules Evolving Reploid (... ∞))
- Runs fully in the browser (offline-capable)
- Self-modifies: tools live in the VFS and hot-reload instantly
- Sandbox focus: IndexedDB filesystem, verification, arena, audit trail
- The cyclical acronym taxonomy is documented in docs/CYCLICAL_ACRONYMS.md
git clone https://github.com/clocksmith/reploid
cd reploid/doppler/reploid
npm install
npm run dev
# Open http://localhost:8080Or use the hosted versions:
- replo.id - Landing page
- replo.id/r - REPLOID agent sandbox
- replo.id/d - DOPPLER inference demo
- Agent receives a goal
- LLM decides which tool to call
- Tool executes against the VFS (hot-reloadable)
- Results feed back to the agent
- Repeat until done (or iterate forever in RSI mode)
- Study recursive self-improvement without host access
- Let the agent edit its own substrate live (tools, prompts, core loop)
- Explore containment: verification, audit, arena gating, HITL
- Self-audit: read the agent loop, find a weakness, patch it, hot-reload
- Recursive Iframes: awaken a child agent in an iframe and report depth
- Ouroboros: improve the code that improves code
- Houdini: probe the browser sandbox for escape vectors
- UI Evolution: iteratively evolve your UI (v1 → v2 → v3...), hot-swapping each version while preserving all interactive controls
- "Escape the box":
Designed for local-first use, but supports frontier models:
| Mode | Provider | Notes |
|---|---|---|
| Local (Browser) | WebLLM | Runs in browser via WebGPU, fully offline |
| Local (Browser) | Transformers.js | ONNX/WASM CPU/Metal fallback, widest compatibility |
| Local (Browser) | DOPPLER | .rdrr format, GGUF import, Native Bridge |
| Local (Server) | Ollama, vLLM | Local server, connect via proxy |
| API | OpenAI, Anthropic, Google, Groq | Direct from client or via proxy |
The proxy (npm start) routes API calls through your machine for CORS and key management.
- VFS: IndexedDB filesystem with genesis snapshots and export/import
- Tool Runner: dynamic tools in
/tools, hot-reload, native tool schemas - Agent Loop: Think → Act → Observe with multi-tool batching and circuit breakers
- Workers/Arena: parallel workers with allowlists; arena for model selection
- Verification/HITL/Audit: optional syntax checks, human approval, and audit logs
- Node.js 18+
- Modern browser: Chrome 90+, Firefox 88+, Safari 14+, Edge 90+
- For local models: 4-8GB RAM, WebGPU-capable GPU recommended
- Quick start guide: docs/QUICK-START.md
- Architecture and boot sequence: docs/SYSTEM_ARCHITECTURE.md
- In-substrate agent manual: doppler/reploid/docs/SUBSTRATE.md
- DOPPLER setup (local WebGPU): docs/DOPPLER_SETUP.md
- Local models (Ollama, WebLLM): docs/LOCAL_MODELS.md
- Security model: docs/SECURITY.md
- Contributing: docs/CONTRIBUTING.md
- Troubleshooting: docs/TROUBLESHOOTING.md
- Showcase runs: doppler/reploid/runs/showcase/
- Small local WebLLM models struggle with tool use and codegen; frontier APIs or stronger local models work best.
- Browser resources (memory/VRAM) bound DOPPLER for large models; prefer MoE or unified memory.
MIT