Browser agents for humans and AI
Give it a website and a task, and get back structured data. Run it here, from your own agent, or in your own Chrome.
Try a sample task
93.9% on WebVoyager
Averaged over two full runs of the 641-task WebVoyager set on live websites (93.1% and 94.7%). TinyFish publishes 91.1% from its own run, with a different harness, patch set and grader.
93.9%
WebVoyager
597 and 607 of 641 tasks passed
91.4%
WebVoyager, strict
Each claim shown in a screenshot
78.3%
Online-Mind2Web
241 and 229 of 300 tasks passed
WebVoyager, 641 tasks
Online-Mind2Web, 300 tasks
TinyFish's numbers are as published; we did not re-run them. Ours are the mean of two full runs of each benchmark. We trail TinyFish on Online-Mind2Web, and the report shows where. Every task result is published.
One tool call from any agent
Add it to Claude Code with one command, or call it from your own code.
Structured output
Pass a JSON Schema and get data back in exactly that shape.
Batches in parallel
Send one URL or a whole list in a single request.
Live progress
Stream every step as it happens, or poll for the result.
Works where you do
MCP for Claude Code and other agents, SDKs for TypeScript and Python, or plain HTTP.
1Install
npx -y @predotdev/mcp setupRun this in your terminal. It signs you in and adds pre.dev to Claude Code and your other coding agents. No Node.js? Use claude mcp add --transport http predev https://api.pre.dev/mcp instead.
2Use it
> Use the browser_agent tool to extract the product title and price from https://www.apple.com/iphone-17-pro/ and return the JSON.Claude Code now has a browser_agent tool. Just ask.
Or run it in your own Chrome
Any coding agent can use the Chrome you already have open, signed in to everything you use.
Your profiles and logins
Every Chrome profile, every open tab and every site you are signed in to: dashboards, internal tools, admin panels.
Works with any coding agent
Part of the pre.dev MCP, so Claude Code, Codex, Cursor or any other MCP client also gets cloud browser agents and specs.
Plain-words actions
Say "click the Create button in the dialog" and it happens in one call, on your pre.dev API key, for a small fraction of a credit.
Free and open source
MIT licensed. Tabs, clicks, typing, reading and screenshots run locally, without calling pre.dev.
Using Claude in Chrome?
Run the setup command, then turn Claude in Chrome off in Claude Code with /chrome. Your Chrome profiles, tabs and logins carry over.
Run this in your terminal
npx -y @predotdev/mcp setupRun it again any time to update.
What it does
- 1Signs you in to pre.dev in your browser, or creates a free account, and saves your key for every agent.
- 2Adds the pre.dev MCP to every coding agent on your Mac: Claude Code, Codex, Cursor, Hermes, Pi, OpenClaw, GitHub Copilot CLI, Gemini CLI, Windsurf, VS Code, Claude Desktop, the pre.dev CLI and more than 15 others.
- 3Connects to your Chrome. Turn on remote debugging once when it asks, then click Allow.
Or let your agent do it
Run `npx -y @predotdev/mcp setup` with a 10 minute timeout and tell me what to click.Paste it into Claude Code, Codex or any other coding agent.
Or use the pre.dev CLI
curl -fsSL https://pre.dev/install | bashThen type /chrome in it. It uses your CLI login.
Needs macOS, Google Chrome and Node.js 22 or later.
Cloud
- Runs in
- pre.dev's browsers
- How
- Send a task and a URL by API, SDK or MCP, get structured data back
- Best for
- Public pages, many runs in parallel
- Setup
- None
Local
- Runs in
- Your own Chrome
- How
- Your agent drives it step by step
- Best for
- Dashboards, internal tools, admin panels, anything behind a login
- Setup
- One command, nothing to host
Start free. Pay per task.
Tasks start at 0.1 credits, about a cent, and more complex tasks can cost more. Credits on every plan work for browser agents, including the free one.