A persistent, daemon-based AI coding agent that runs fully on-device (Android via Termux or Linux) with zero cloud dependency by default.
Works out of the box with:
- Qwen2.5-Coder-7B (main coding/reasoning agent)
- Qwen2.5-0.5B (planner + summarizer)
- nomic-embed-text-v1.5 (RAG embeddings)
All served locally via llama.cpp.
Super easy OpenRouter setup (cloud fallback or free models):
- Get your key at openrouter.ai/keys (starts with
sk-or-) - Set:
export OPENROUTER_API_KEY="sk-or-your-key-here"
export CODEY_BACKEND="openrouter"
Add to ~/.bashrc and source ~/.bashrc
- That's it! Defaults to
qwen/qwen-2.5-coder-7b-instructfor coding + planner. No local models, no llama-server, no daemon needed.
Optional: Override with OPENROUTER_MODEL and OPENROUTER_PLANNER_MODEL (e.g. free or powerful models from OpenRouter).
Just pip install -r requirements.txt and run python main.py "your task".
Highlights:
- RAG-powered local knowledge base
- Recursive self-refinement (draft β critique β refine)
- Smart error recovery & adaptive tools
- Git integration (branches, AI commits, conflict resolution)
- Voice support (TTS/STT on Termux)
- Thermal & resource awareness (great for mobile)
- Auto Python linting + /review command
- Persistent state across sessions
- Fine-tuning ready (export history for Colab)
Full privacy on-device or quick switch to OpenRouter. One-command local install with ./install.sh + codeyd2 start.
Check it out and start coding!
π https://github.com/Ishabdullah/Codey-v2
Top comments (0)