Jev MCP (Browser Automation)

A Model Context Protocol server that lets Claude Code, Codex and any MCP-capable agent drive a real Chrome browser while JEV — TypeSafe's System One decision model — answers typed questions like which element is the departure-city input, in about 100 ms each.

They connect as MCP clients; JEV makes the many small decisions so your agent stays the planner. Calibrated confidence drives act, verify or ask: the tool acts on its own when it is sure and asks your agent when it is not. Every decision is traced and can be replayed.

The README's table: JEV decisions take about 100 ms against 2–20 s per step for frontier LLM agents; output is typed choice, score or yes-no instead of free text to parse; cost is $0.042 per million input tokens with output free; and your agent's context stays compact instead of filling with HTML and screenshots.

A local airline-search imitation with Russian UI, a consent wall, autocomplete and a low-fare calendar: finding the cheapest Almaty to Antalya flight in October took 7 steps, 20 JEV calls, 0 questions to the agent, $0.0015 and 16 seconds, ending with the site's own cheapest-first sorting.

Its grounding eval ran 58 labelled cases over 10 fixture sites — distractor fields, unlabeled inputs, shadow DOM, cross-origin iframes and not-on-this-page cases — reporting 100% top-1 accuracy, a median decision time of 0.44 s and expected calibration error of 0.05. These are the project's own fixtures, not independent tests.

Your agent plans; JEV executes the decisions. The README is explicit that JEV does not write text or plan — it turns the page into a clean model, and JEV decides which element, which autocomplete suggestion, whether a cookie wall blocks the page and whether a step succeeded.

The README badges Node 22.18+ and an MIT license, and its comparison table prices the JEV API at $0.042 per million input tokens, so expect a TypeSafe key and a local Chrome. The repository README is longer than the excerpt available here, so check the project website linked from the repo for exact setup steps.