Storage1n1

Local Computer Use Agent — Development skill for Claude Code

Development community

Free, fully offline computer-use agent (Set-of-Marks grounding + crop-zoom super-resolution) that drives a real desktop with small local VLMs via Ollama - a local alternative to Claude Computer Use, O.

How to install Local Computer Use Agent

This entry records only its repository, not the path inside it, so there is no exact command to give. Open Storage1n1/local-computer-use-agent and copy the folder into ~/.claude/skills/, or the file into ~/.claude/agents/.

What Local Computer Use Agent does

Alternatives in Development

Internet Court Skill — The trust layer for agent-to-agent commerce — natural-language mandates, ERC-7710 delegated permissions, x402 4.9k ★
08 Open Question Resolution — Open Question Resolution 1.1k ★
Notebook Intelligence — A JupyterLab extension supporting Claude Code, Copilot, Ollama, and OpenAI-compatible LLMs, with MCP, skills 331 ★

README

Local Computer-Use Agent

**Architect: SAKSHAM SWAMI**

A free, fully **offline** computer-use agent: a local vision-language model (via [Ollama](https://ollama.com) — `qwen3-vl`, `minicpm-v4.6`) watches the screen and drives real mouse/keyboard actions to complete a task, end to end, with no cloud API and no per-action cost.

This is the same category of system as **Anthropic's Claude Computer Use**, **OpenAI's Operator / ChatGPT Atlas**, and **Google's Project Mariner** — an LLM that can see a screen and operate a real desktop or browser — except it runs entirely on your own machine against small, local, free models instead of a hosted frontier model. The bet here: small local VLMs are bad at raw pixel-coordinate prediction, but that's a solvable *grounding* problem, not a model-size problem — so this project focuses almost entirely on the scaffolding that makes a small model's job trivial instead of trying to make the model itself smarter.

See it work

Filling a form	Drawing a house

Full screen recordings of both runs, **fully autonomous, unedited**:

`videos/form_filling_demo.mp4` — launches a Tkinter registration form from a bare desktop, grounds and fills Name / Email / City, checks the terms checkbox, and submits.
`videos/drawing_demo.mp4` — launches GNOME Drawing, focuses the canvas, and draws a house (roof, walls, door, window) via smooth interpolated drag strokes.

The form run's exact output, written by the agent's own click on Submit: [`docs/example_form_submission.json`](docs/example_form_submission.json).

Why small local models need help that big API models don't

Frontier computer-use models (Claude, GPT) are trained specifically to predict screen coordinates directly. Small local models (`qwen3-vl`, `minicpm-v4.6`, `gemma4` — the kind that run on a single consum