Local-first task automation using a fine-tuned 0.5B LLM. No cloud APIs, no subscriptions, runs entirely on your CPU.
ACE detects task types automatically and generates execution plans - all running locally with a custom-trained 0.5B language model optimized for CPU inference.
# Install
git clone https://github.com/ansh0x/ace
cd ace
pip install .
# Initialize (downloads models)
ace init
# Run tasks
ace run "copy /home/user/logs/*.log to /backup/"
ace run "open reddit.com"
ace run "extract emails from contacts.txt"ACE takes natural language tasks and:
- Detects task type (atomic, repetitive, or needs clarification)
- Generates execution plan (CLI commands + hotkeys)
- Executes safely (shows plan, optionally auto-executes)
All processing happens locally - your data never leaves your machine.
✅ Local execution - No cloud APIs, no data transmission
✅ Task type detection - Automatic identification of atomic vs repetitive tasks
✅ CPU-optimized - Runs on modern CPUs without GPU (3-10 sec on i3/i5)
✅ Privacy-first - All processing happens on your machine
✅ Quantized models - Efficient inference with GGUF Q4 quantization
| Hardware | Performance |
|---|---|
| Intel i5 (2018+) / Ryzen 5 + SSD | 3-5 seconds per task |
| Intel i3 (2015+) + SSD | 5-10 seconds per task |
| Older hardware / HDD | 30-90 seconds per task |
See ROADMAP for planned improvements.
- Python 3.8+
- 8GB RAM minimum
- 2GB disk space for models
git clone https://github.com/yourname/ace
cd ace
pip install .
ace init # Downloads models to ~/.ace/If ace init fails, download models manually:
- Task models: HuggingFace
- Place in
~/.ace/models/llm/
# Run a task
ace run "your task here"
# Use quantized model (faster, slightly lower quality)
ace run -q "your task"
# Disable caching (slower but fresh results)
ace run --no-cache "your task"
# Verbose output (for debugging)
ace run --verbose "your task"ace hotkey list
ace hotkey list -g # List groupsConfiguration file is in ~/.ace/config.json
- Right now there no customizations available.
- They will be avaialble in future
- Base Model: Qwen2-0.5B fine-tuned with LoRA
- Training Data: ~1000 task examples (atomic + repetitive)
- Quantization: GGUF Q4_K_M (300MB)
- Inference: llama.cpp (CPU-optimized)
- Search: Semantic hotkey/file matching with sentence-transformers
Trained on custom dataset covering:
- File operations (copy, move, delete)
- Browser automation (for now only opens browser and websites, refer to example/youtube_test.mkv)
- Task classification (atomic vs repetitive vs clarification)
v0.2 (In about a month)
- Smart file path detection
- Performance optimizations (special tokens)
- Better error handling
- More customizable using the config.json
v0.3+
- VLM integration for visual tasks
- YOLO for better UI navigation
- Improved repetitive task handling
- User corrections → model improvement
Contributions welcome! Areas that need help:
- Testing on different hardware
- Testing workflow on Windows
- Documentation improvements
- Bug reports and feature requests
Please open an issue before starting major work.
AGPL-3.0 - See LICENSE
This means:
- ✅ Free to use, modify, distribute
- ✅ Must share source code of any modifications
- ✅ Must use AGPL-3.0 for derivative works
- ❌ Cannot use in closed-source commercial products without permission
CC BY-NC-SA 4.0 - See model repository
- ✅ Free for personal and research use
- ❌ Commercial use requires separate license
- ✅ Must credit the author
- ✅ Derivatives must use same license
Training of AI/ML models using this code or model weights is prohibited without explicit written permission.
For commercial licensing inquiries: [your email]
Built with: PyTorch • Transformers • llama.cpp • sentence-transformers
Models: HuggingFace | Discuss: Issues
