OpenPC
The Local-First
Autonomous Computer Operator
OpenPC is a Python-based autonomous AI agent that runs entirely on your local machine. It uses a Synchronous Lock-Step State Machine to perceive your screen, plan complex tasks, and execute actions—without ever sending data to the cloud.
Our Mission: Democratize AI Autonomy
We believe autonomous AI should be accessible, private, and user-controlled. OpenPC challenges the cloud-dominated landscape by proving that sophisticated AI agents can run entirely on local hardware, giving users complete ownership of their digital workflows.
Research Foundations
- Lock-Step State Machine Architecture — Synchronous perception, planning, and execution cycles prevent race conditions and ensure deterministic behavior
- Coordinate-Free Visual Grounding — GUI-Actor's heatmap approach eliminates brittle DOM parsing and UI element dependencies
- Multi-Model Collaboration — Specialized models for vision, planning, verification, and clarification create robust decision-making pipelines
- Local-First Design Philosophy — Zero telemetry, complete data privacy, and offline operation as core architectural principles
Impact & Vision
By 2025, we aim to make local AI autonomy the standard for personal computing, enabling anyone to automate their digital life without compromising privacy or control.
Zero Telemetry. Unlimited Potential.
100% Local
Vision via GUI-Actor and Hugging Face. Planning via Qwen2.5 and Llama-3.2 through Ollama. Zero API tokens required.
Full Autonomy
No hardcoded shortcuts. The planner decides every step from the goal and the screen context using its own intelligence.
Dual Execution
Standard GUI mode using PyAutoGUI or "God Mode" utilizing direct shell and HTTP commands for blazing fast navigation.
Self-Correcting
Failed actions are logged and never repeated. Stuck detection forces strategy changes autonomously.
Shape the Future of Autonomy
AI Core Researcher
Optimize vision grounding and strategic reasoning loops. Work directly with GUI-Actor and Qwen2.5 fine-tuning.
View Details →Systems Engineer
Master of GPU offloading and local state orchestration. help us squeeze every millisecond out of consumer hardware.
View Details →Product Designer
Craft the next generation of human-AI interfaces. Minimalist, high-performance, and deeply intuitive design.
View Details →Want to see all open roles?
Explore Full Careers PagePurpose-Built Local Intelligence
GUI-Actor-7B
Chosen for its best-in-class, coordinate-free heatmap grounding. It maps screenshots and goals to precise screen pixels instantly, avoiding the overhead of heavy DOM manipulation or UI element trees.
Qwen2.5 14B
Selected for exceptional strategic reasoning capabilities. It is the perfect engine for reading associative memory, evaluating vision outputs, and picking precise atomic actions with complete autonomy.
LLaVA 7B
Integrated for fast, reliable visual validation. Highly optimized for before/after screenshot comparisons, instantly catching wrong-UI interactions and providing actionable failure logs to the planner.
Llama 3.2 3B
Utilized as an ultra-lightweight disambiguation gate. Extremely efficient at pre-loop analysis, asking clarifying questions only when a goal lacks clear deterministic intent.
Built for Enthusiast Hardware
GPU Constraints
An NVIDIA GPU with 16 GB+ VRAM is strongly recommended (system tested extensively on RTX 4060 Ti 16GB). OpenPC forces all inferences onto the GPU for rapid, lock-step processing.
Local Stack Prerequisites
Requires Ollama running locally (with num_gpu bypassed) to host the Planner, Verifier, and Clarifier. Vision is handled through a Hugging Face PyTorch installation.