Give your LLM a browser and a budget

What it does

Open Browser is a TypeScript monorepo that wraps Playwright in an agent loop. You describe a web task in plain English, and the framework repeatedly snapshots the page, feeds the DOM state and your goal to an LLM, and executes the resulting commands—click, type, navigate, extract—until the job is done or it hits a limit. It ships as a core library, a CLI, and a sandboxed runtime with resource guards.

The interesting bit

The framework treats the LLM as a literal remote operator. It includes an interactive REPL where you can issue browser commands manually, plus built-in metering for API spend and a sandbox that enforces CPU, memory, and domain restrictions. That combination of “let it roam free” and “here is your leash” is what the README sells as production-ready.

Key highlights

• Swaps between OpenAI, Anthropic, and Google models via a single flag through the Vercel AI SDK.
• Interactive browser> REPL for debugging and prototyping live sessions.
• Sandboxed execution with timeouts, memory caps, and allowed-domain lists.
• Built-in stall detection, cost tracking, session replay recording, and failure thresholds.
• Exposes both high-level Agent API and low-level browser commands (screenshot, eval, extract) for direct control.

Verdict

Worth a look if you are building autonomous agents that need to interact with real websites and want guardrails around spend and scope. Skip it if you just need static scraping—Playwright alone is lighter.