jev-browse

jev-browse drives a real browser with TypeSafe Jev making every decision and agent-browser performing every action. You give it a goal, such as open the Hacker News FAQ. Jev picks each click, chooses which of your texts to type into which box, and decides when the goal is reached or can't be reached.

jev-browse is an unofficial project and isn't affiliated with TypeSafe or Vercel.

Benchmark

The following table compares jev-browse with Claude Code driving agent-browser directly on the same 10 tasks, 3 runs each. Claude Code ran on 16 September 2026, and jev-browse ran on 17 September 2026 with its default network wait. Both agents had the same limits: agent-browser commands only, no entering URLs, and only the texts given for the task. Every run started with cleared cookies and site storage.

Agent	Passed	Median time per task	Cost per task
jev-browse (Jev + agent-browser)	30/30	4.4 s	$0.0009
Claude Code (Claude Sonnet 5) + agent-browser	30/30	9.4 s	$0.0679

The tasks include navigation goals on Wikipedia, Hacker News, example.com, and the TypeSafe docs, 2 goals that need typing, and 2 goals that the agent must refuse, such as logging in without credentials. The longest task runs a checkout on the Sauce Labs demo shop: log in, add 2 items to the cart, fill in the checkout form, and stop before Finish. jev-browse finished it in 23.3 to 23.5 s for $0.0010. Claude Code took 19.8 to 29.3 s for about $0.08.

The results show that jev-browse has a lower median time and costs about 70 times less on these tasks. On the checkout task, the default network wait makes jev-browse about as fast as Claude Code. They don't show how it performs on longer or more complex tasks.

Jev cost is the input tokens at $0.042 per million tokens. Claude Code cost is the total_cost_usd value that Claude Code reports at API list prices. To rerun the benchmark, see Run the benchmark.

How it works

Each step sends one request to Jev, which answers 3 kinds of questions at once:

• Is the current page the goal?
• Does the goal need an action that the agent can't do?
• Which action makes the most progress: a click, typing one of your texts, pressing Enter, or going back?

Jev answers with probabilities and never writes text. Any search query or form value comes from you, or from the AI agent that calls the MCP server.

Limits

jev-browse has these limits:

• It types only the texts that you pass. It can't write its own text or enter URLs.
• It logs in only with a username and password that you pass as texts. It can't create accounts, solve CAPTCHAs, hover, drag, upload files, or download files.
• When several elements on a page have the same label, such as a row of Add to cart buttons, it sees only the first one.
• It doesn't write answers or summaries. A run succeeds when it ends on the page that the goal names. With --extract, it returns matching page lines verbatim, and it can miss some of them.
• It sends the page content and your texts, including any password, to TypeSafe with every request. Pass only test or throwaway credentials.

Requirements

You need the following:

• Node.js 22.9 or later.
• agent-browser, installed so that agent-browser --version works.
• A TypeSafe API key in the TYPESAFE_API_KEY environment variable.

Use the command-line tool

To reach a page, run the following command:

npx github:kyrylosyzonenko/jev-browse "open the Hacker News FAQ" https://news.ycombinator.com

To let the agent type, pass each text with --text NAME=VALUE:

npx github:kyrylosyzonenko/jev-browse "open the Wikipedia article about Kermit the Frog" https://en.wikipedia.org/wiki/Rubber_duck --text "query=Kermit the Frog"

The command-line tool has these options:

Option	Default	Description
--text NAME=VALUE	None	A text that the agent may type. Repeat the option for more texts.
--steps N	15	The maximum number of steps.
--done P	0.8	The probability at which a run stops as reached or blocked.
--extract WHAT	None	Describes the data to return from the final page, such as hotels, each with its price. When the goal is reached, Jev selects the page lines that match, and the command prints them verbatim after extracted.
--no-network-wait	Off	Skips waiting for network activity to end after every action. Runs are faster, but sites that reload in the background, such as Google Hotels, can reset typed values.

The command exits with one of these codes:

Code	Meaning
0	The goal was reached.
1	The agent found no useful action or hit the step limit.
2	The command was used incorrectly.
3	The goal needs an action that the agent can't do.
4	The run crashed, for example because the API key is missing.

Use the MCP server

The MCP server exposes one tool, browse. The tool takes a goal, an optional url, optional texts such as {"query": "Kermit the Frog"}, an optional steps limit, an optional extract description of the data to return, and an optional networkWait switch that is true by default. It returns a status of reached, stopped, blocked, or error, the final URL, the step log, and any extracted lines. The browser stays open afterward, so the calling agent can read the final page with agent-browser.

To add the server to Claude Code, run the following command:

claude mcp add jev-browse -e TYPESAFE_API_KEY=YOUR_API_KEY -- npx -y -p github:kyrylosyzonenko/jev-browse jev-browse-mcp

To add the server to Codex, run the following command:

codex mcp add jev-browse --env TYPESAFE_API_KEY=YOUR_API_KEY -- npx -y -p github:kyrylosyzonenko/jev-browse jev-browse-mcp

To add the server to Hermes Agent, run the following command:

hermes mcp add jev-browse --command npx --env TYPESAFE_API_KEY=YOUR_API_KEY --args -y -p github:kyrylosyzonenko/jev-browse jev-browse-mcp

Any other MCP client that runs stdio servers can use the same npx command.

Run the benchmark

To run the benchmark from a clone, follow these steps:

1. Clone the repository and install the dependencies:

   git clone https://github.com/kyrylosyzonenko/jev-browse.git && cd jev-browse && npm install
   

2. Copy .env.example to .env, and put your TypeSafe API key in it.

3. Run every task 3 times with both agents:

   node --env-file-if-exists=.env eval.mjs 3
   

The Claude Code runs need the claude command and cost real money, up to $2 per run. To change the Claude model, add --model opus. To run only one agent, add --only jev or --only claude.

Support me

I build jev-browse in my own time, and it's free to use. If you want to support me, you can send a tip on Ko-fi.

License

jev-browse is released under the MIT License.