A fully autonomous, local-only coding agent capable of auto-evolving its own toolset without breaking its execution loop.
This is a personal learning project created for educational purposes and to explore different code concepts with ai concepts.
- Status: Personal sandbox / Portfolio piece.
- License: This project is open-source and available for public educational use under the MIT License.
- Purpose: Academic research and technical skill development.
Build a fully autonomous, local-only coding agent capable of auto-evolving its own toolset without breaking its execution loop. Must operate strictly within 12GB VRAM limits (targeting ~50 t/s) using Gemma-4 E4B/E2B (Multimodal) via local LM Studio.
- Frontend UI/UX: React (TypeScript), Tailwind CSS, shadcn/ui.
- Frontend State: Zustand.
- Backend (Core): Python (FastAPI).
- Data Validation: Pydantic (Backend) / Zod (Frontend).
- AI Orchestration: Hybrid Stack (PydanticAI for Planning/Validation, SmolAgents for Execution).
- API Proxy: LiteLLM (bridges OpenAI schema to target models).
- Local LLM Provider: LM Studio (http://127.0.0.1:1234/v1).
- Language: TypeScript (Node/Bun) or Python.
- JIT Context Engine (Tree-Sitter): Parses the workspace using Tree-sitter on startup/file-change.
- Guardrails & File Sandbox: Jails all execution to the
./workspacedirectory. Safe deletes route to./workspace/.trash/. - The Reflexion Loop (Auto-Retry): Catches malformed JSON or tool execution errors and feeds them back to the LLM.
- Multimodal LLM Handler (Gemma-4): Supports base64 image ingestion and streams
<|think|>tokens to a separate UI stream. - Auto-Evolution Engine (Tool Factory): AI writes new tool scripts + JSON schemas with automated validation.
./
├── .agents/ # Core configuration and static rules
├── workspace/ # The AI's sandbox - ALL AI work happens here
│ ├── .trash/ # Where "deleted" files go
│ └── .agents/
│ └── dynamic_tools/ # AI-evolved toolsets
├── core/ # The Immutable Core (FastAPI Backend)
│ ├── main.py # Entry point
│ ├── api/ # API Endpoints for UI
│ ├── services/ # Business logic
│ ├── modules/ # Core subsystems
│ ├── schemas/ # Pydantic models
│ └── tests/ # TDD test suites
└── ui/ # Observer Dashboard (React + Vite)
For more detailed information, please refer to ARCHITECTURE.md and FILE_STRUCTURE.md.