fx-goal-browser

An experiment in fast, goal-driven browser use. Give it an objective in plain language; get a structured JSON answer.

1. Set the goal. fx turns your request into a clear objective and decides what information counts as success.
2. Define the result. Before browsing, the agent creates a Zod schema: a contract for the answer's fields, types, and constraints.
3. Browse fast. Jev follows observe → choose a reference → act, using accessibility snapshots to select page elements. One persistent Playwright browser keeps actions fast. It works toward the goal across pages and steps.
4. Extract the answer. Once the information is available, Luna reads the recorded snapshots and returns the answer in one structured-output call, with the goal's Zod schema supplied directly to that call.

All AI calls go through Vercel AI Gateway. A URL in your prompt is used directly; otherwise, agentic web search finds a starting page.

Try it

Requires Node.js 24+. Uses TypeScript 7.

npm ci
npm run browser:install
npm run build
cp .env.example .env

Set AI_GATEWAY_API_KEY in .env, then run:

node dist/cli.js --headed "Find the top 5 movies currently trending on Rotten Tomatoes to watch at home. Rank them by Popcornmeter score, highest first. Return each movie's title, score, and URL."

The result goes to stdout as JSON; progress goes to stderr. There is no default step limit. Ctrl+C cancels. Use --help for model overrides and other options.

Run checks with npm run check && npm test. Measure local browser speed with npm run bench. Tests use fictional fixtures and make no paid AI calls.

Credentials and browsing records stay in ignored .env and runs/ files.