Jev Browser Control

Let Claude control your own Chrome. A Chrome extension plus a small MCP server: Claude plans, and Jev, TypeSafe's decision model, picks each click, keystroke and scroll in about half a second for a fraction of a cent.

Created by Jeroen Erne (nexibeo.com · completeaitraining.com), built together with Claude.

Measured 27 to 198 times cheaper, and more than twice as fast, than letting Claude's or Codex's own model do the clicking (same tasks, same loop; results).

Website: jevbrowsercontrol.com · Docs: jevbrowsercontrol.com/docs · License: MIT

What you get

• A Chrome extension that reads the page as numbered elements and acts on them with trusted input. Run tasks from its side panel, or let Claude drive it.

• An MCP server for Claude Code and Claude Desktop with 17 tools: browser_snapshot, browser_click, browser_type, browser_navigate, browser_read, browser_screenshot and more, plus jev_task (hand a whole sub-task to Jev), jev_find and jev_check.

• Your choice of who pays for Jev: your own OpenRouter key (free, this repo), prepaid credits from jevbrowsercontrol.com (one key, 5× OpenRouter's price), or any compatible endpoint such as TypeSafe direct.

Related MCP server: Browser Tools for Claude Code

Quick start

1. Install the extension. Clone this repo (or download the zip), open chrome://extensions, switch on Developer mode, click Load unpacked and pick the extension/ folder.

2. Choose a provider in the settings page that opens: paste an OpenRouter key or a jbc_ credits key, and press Test connection.

3. Connect Claude. Needs Node.js 18+; no other dependencies.

   claude mcp add jev-browser -- node /path/to/jev-browser-control/mcp/server.mjs

   or without cloning:

   claude mcp add jev-browser -- npx -y https://jevbrowsercontrol.com/downloads/jev-browser-control-mcp-0.2.0.tgz

   Claude Desktop: add {"mcpServers": {"jev-browser": {"command": "node", "args": ["/path/to/mcp/server.mjs"]}}} to its config.

Then ask Claude: “Use the browser to search Wikipedia for Ristretto and tell me where the name comes from.”

Grok Bot, ChatGPT, claude.ai and other cloud apps. They can't start a local program, so jevbrowsercontrol.com offers the same tools as a remote MCP server at https://jevbrowsercontrol.com/mcp (bearer: your jbc_ key). Switch on Remote AI apps in the extension's settings and the extension keeps an outbound connection to the relay; it's off by default, and clicks that buy, pay, send, post or delete always wait for your OK in the side panel. For Grok Bot there is a ready-made template: Jev Browser Operator.

Claude Code and OpenAI Codex agent files. agents/ has a skill (works in both) and a Claude Code subagent. node scripts/install-agents.mjs registers the MCP server and installs them for whichever of the two you have.

How a step works

page ──► numbered elements ──► one Jev call ──► stop gates ──► trusted input ──► page
         (extension, ~20 ms)   (~0.5 s)         (code)          (CDP, ~50 ms)

1. Snapshot. Visible buttons, links and fields become rows [n] role "label" value, with checked/selected state and nearby card text. Password and file fields are never listed. Open shadow roots are included.

2. One request, three answers. Jev is asked for the operation (CLICK, TYPE_TEXT, SELECT, PRESS_ENTER, SCROLL_*, BACK, WAIT, DONE, BLOCKED), speculatively for the best target of every operation, and, independently, whether the goal is already met. Code uses only the target that matches the chosen operation.

3. Stop gates in code. DONE counts only when the independent goal check agrees; clicks that buy, pay, send, post or delete stop for confirmation; three actions without a visible change end the run; limits on actions, seconds and dollars.

4. Act. Before input, the element is re-checked (same node, same state, visible, not covered). If the page moved while Jev decided, nothing is clicked and it looks again. Only TYPE_TEXT needs words: a small LLM writes them from the goal and never invents personal data (it returns null, and the task stops with needs_input).

Jev's answer is always an element number that code maps back to a node it tagged itself. It never becomes a selector, coordinates or code.

The operation-plus-speculative-target design, the freshness guards and the page snapshot are adapted from Browser Use's Jev Ultrafast (MIT); see NOTICE.

Measured

Live runs on September 19, 2026 with typesafe/jev-1.13 (cost at OpenRouter prices; credits cost 5×):

Recorded traces are in results/.

Jev vs. Claude and Codex models

scripts/benchmark.mjs runs three live tasks (search Wikipedia and open an article; fill 7 fields of httpbin's order form without submitting; open the top Hacker News story's comments) through the same loop and page snapshot, changing only who answers each step's questions. With --session, the models run the way Claude Code and Codex run them: tool definitions in the system prompt, the whole conversation resent at every step, reasoning at medium effort, prompt caching on. Every model passed all three tasks; success is checked in code on the final page.

September 19, 2026, OpenRouter prices. Without --session (one bare call per step, reasoning off or low) the models cost about the same or less: $0.167, $0.247, $0.538 and $0.941. Prompt caching keeps the resent conversation cheap on short tasks like these; a long real session, which carries all its earlier work into every step, costs more. On Hacker News Jev reached the right page each time but its own goal check was unsure, so it ended as stuck or done_unconfirmed rather than done. Raw data: results/benchmark-session-2026-09-19.json and results/benchmark-2026-09-19.json.

Repository

npm install                 # playwright-core, only for e2e and icons
npm test                    # 20 offline tests
CHROME_PATH=... OPENROUTER_API_KEY=... npm run e2e    # live, about $0.005
npm run zip                 # dist/ extension zip + MCP tarball

CHROME_PATH must be Chromium or Chrome for Testing: branded Chrome 137+ ignores --load-extension.

MCP server settings

Safety and limits

• Remote control is off until you switch it on. While it's on, anyone with that jbc_ key can reach your browser through the relay, so keep the key secret and revoke it in the dashboard if it leaks. Remote callers can never skip the confirmation for irreversible clicks.

• The bridge listens on 127.0.0.1 only. Web pages can't connect (they can't forge a chrome-extension:// origin); other MCP sessions need a token from your home folder.

• The agent acts in your normal profile, with your logins. Use a separate Chrome profile for risky work, and list sites like your bank under Blocked sites.

• Page text is sent to Jev as untrusted data, but prompt injection can still mislead a model. Keep limits on and check results; Jev can pick a confident near-miss between look-alike names.

• Not supported yet: cross-origin iframes, canvas apps, file uploads, CAPTCHAs, and hover-only menus. English pages work best.

Credits

Created by Jeroen Erne (nexibeo.com · completeaitraining.com), built together with Claude. Loop design from Jev Ultrafast by Browser Use. Jev is a model by TypeSafe. This project is not affiliated with TypeSafe, Browser Use or Anthropic.