FreeComputerUse

Quick start · Project site · Try the task lab · Security

Tell your browser what done looks like. FreeComputerUse handles the steps, while you keep control.

FreeComputerUse is a local-first AI browser automation agent. A model plans small action batches; TypeScript and Playwright execute and verify them in your browser. Compatible workflows can be learned once and replayed later with zero model calls.

Watch a real run

One goal led through six pages to a verified incident brief: 21 successful actions, 10 model calls, one repaired failure. The configured cost estimate was $0.00650. This is one recorded synthetic task, not a general success-rate claim. Watch the full recording · Inspect the evidence.

How it works

1. Observe the page’s DOM and accessible controls.
2. Plan a bounded batch with the model you choose.
3. Execute browser actions locally with Playwright; approve each new website origin first.
4. Verify results, repair only what failed, then reuse compatible learned workflows.

The model does not run shell commands or arbitrary JavaScript. Browser previews stay local; the goal and selected page context are sent to your configured model provider.

Quick start

Requires Node.js 22.13+, npm and Playwright Chromium.

git clone https://github.com/OthmaneBlial/FreeComputerUse.git
cd FreeComputerUse
npm ci
npx playwright install chromium
cp .env.example .env
chmod 600 .env

Set a model API key in .env, then run the local workspace:

LLM_PROVIDER=openai-compatible
LLM_API_KEY=your-key
LLM_BASE_URL=https://api.deepseek.com
LLM_MODEL=deepseek-flash

npm run dev

Open http://127.0.0.1:4318, enter a starting URL and goal, then approve site access.

Try it without an API key: open the practice revenue table and run the goal Extract the table. This narrow workflow has a deterministic local strategy; general tasks need a model provider.

Pick your model

Use a provider API key, or sign in through the official CLI for a supported subscription. API usage and consumer subscriptions are separate billing products.

The provider API routes have not all been live-tested here; DeepSeek Flash is the measured default. Subscription modes use the local CLI sign-in, need no API key and respect plan limits. Their local agent tools and MCP servers are disabled while planning. See .env.example for optional model, CLI path and task-budget settings.

Control and privacy

• Website access is approved per origin; sensitive actions have a separate confirmation gate. Ultra mode is explicit and off by default.
• Browser execution, profiles, history and downloads stay on your machine. Page context needed for a plan goes to the chosen model provider; screenshots are not sent.
• The planner receives aliases for local profile and file values, not their contents. Local storage is not encrypted.
• Runs record actions, checks, repairs, token estimates and workflow reuse so you can inspect what happened.

Read the security boundaries before using personal or sensitive data.

Evidence

In a focused public suite recorded 18 September 2026, 14/14 first-run results and 14/14 compatible learned repeats passed independent checks. First runs used 18 model calls; repeats used zero. These single-trial sandbox results do not predict success on arbitrary websites. Report · Method and limitations.

Explore and contribute

• Project site and task library
• Useful browser-task examples
• Implementation notes
• Local validation commands

Useful contributions: reproducible browser tasks, safer permission scopes and checks that make results easier to trust. Keep shared examples synthetic or read-only; remove credentials and personal data from traces.

MIT License · Built with TypeScript and Playwright.