Overview
jev-mcp is a browser automation MCP server powered by JEV, TypeSafe's System One model. It allows MCP clients such as Claude Code and Codex to drive a real Chrome browser. JEV makes typed decisions in approximately 100 ms with calibrated confidence, and asks the agent only when it is unsure. The server provides autonomous task execution, page understanding, human-in-the-loop questioning, and safety gates. It can use either a paired Chrome extension for the user's logged-in browser or its own headful/headless Chrome profile. The system includes a debug UI with probability bars, replay, and calibration tools. Measured results on an airline-search fixture showed a task completed in 8 steps, 24 JEV calls, 0 questions, $0.0019, and 19 seconds, with a grounding eval score of 58/58 (100%) and expected calibration error of 0.05.
Pricing
Free
Detailed plans have not been confirmed in our catalog. Check the official website for current limits and billing terms.
Visit WebsitePrices and limits may change. Confirm the currency, billing period, seat minimum and usage caps on the official website.
Use Cases
- Teams use jev-mcp to automate browser tasks such as dismissing consent walls, filling autocompletes and date pickers, submitting searches, and extracting results.
- Developers use jev-mcp to understand page structure by converting DOM, accessibility tree, and paint order into regions and stable element refs.
- Agents use jev-mcp to ask for human input when decisions have low confidence, receiving questions with candidates and probabilities.
- Users can employ jev-mcp with their own logged-in Chrome browser via a paired Chrome extension.
- Teams use jev-mcp to enforce safety gates that ask before irreversible actions and prevent secrets from reaching the model.
- Developers use the debug UI to inspect probability bars for every decision, replay sessions, and access calibration tools.
Features
Controls a real Chrome browser
Executes multistep browser tasks
Understands page elements and structure
Requests help for uncertain decisions
Gates irreversible browser actions
Protects secrets from model access
Replays decisions in a debug interface
Reviews
No reviews yet. Be the first to share your experience!
FAQ
What is JEV?
JEV is TypeSafe AI's first System One model. It returns typed decisions (choice, score, yes/no) with calibrated probabilities instead of generating text.
How is it different from Playwright MCP or browser-use?
Those tools route every decision through an LLM. jev-mcp moves grounding, assessment and verification to JEV, which is much cheaper and faster and has calibrated confidence. Your agent still does the planning and handles ambiguity.
Which agents can use it?
Any MCP client: Claude Code (with optional channel push), Codex, Cursor and others over stdio.
Can it use my logged-in Chrome?
Yes, through the Chrome extension after a one-time pairing. Without the extension it uses its own profile.
How much does it cost?
jev-mcp is free (MIT). JEV costs $0.042 per million input tokens and output is free; a typical task costs a fraction of a cent.
Is it safe to let it click?
Payments, orders, messages and deletions always ask first, secrets never reach the model, and every decision is traced locally.