jev-browser

Browser automation where an LLM plans and Jev decides.

Unofficial project, not affiliated with TypeSafe. It calls the TypeSafe System One API with your own API key.

The calling LLM (Claude, via MCP) says what outcome it wants, one step at a time, and hands over any text to type. For each round of a step, code describes the page. Then one ~300 ms Typesafe System One request asks Jev several questions at once: which element, which action, which value, and is the step done / blocked / showing an error / about to do something irreversible. Playwright performs the action. The LLM never reads page snapshots unless it chooses to take over.

Claude ── browser_do("Log in", {email, password}) ──▶ jev-browser
                                                       │  loop until done / stuck / needs confirmation
                                                       │   1. settle   (network + DOM quiet)
                                                       │   2. describe (elements, labels, state, visible text, diff, counts)
                                                       │   3. Jev      (done? error? irreversible? tool? target? value?)
                                                       │   4. act      (Playwright)
Claude ◀── { status: "done", url, actions[], done_score } ─┘

With WebMCP, a page offers its own tools. Jev chooses between those and the page's buttons, and fills each argument by picking one of the caller's values (text: values.g1 → "milk"), never by writing text:

Jev only answers with probability distributions: yes/no (noul), pick one option (choice) or a rating (score). It never writes text. So everything free-form comes from the caller as candidates, and code turns disagreement or low confidence into a status the LLM can act on.

42 tasks in 16 categories on live sites (see RESULTS.md):

• 40/42 correct in the latest run, 0 false "done" claims (38/41 twice before the latest fixes). Remaining misses: counting ("add until 3", flagged likely_done) and verifying a sort.
• ~300 ms per Jev call, 2–4 calls for most steps; a 5-step checkout takes ~14 s end to end.
• Pause before irreversible actions: across ~200 rounds it flagged only saucedemo's "Finish" (place order) button.
• Less page content for the LLM than a Playwright-MCP-style loop on the same tasks (estimate): median 5× per task, ≈8k vs ≈557k tokens in total. Big pages dominate the total: a Wikipedia article is ~149k snapshot tokens. On tiny pages there is no saving. The page reading moves to Jev.

Works on: forms, native and custom dropdowns, checkboxes and radios (including styled replacements), dynamic loading, modals, JS dialogs, hover, right-click, drag and drop, key presses, file upload, iframes, shadow DOM, new tabs, pages with 2,000+ elements (two-stage selection), and non-English UIs.

Known limits, so write steps around them:

• Ordered sub-goals in one step ("add two todos, complete one, clear completed") → split into one outcome per step.
• Open-ended goals ("scroll to load more") → make them measurable or check yourself.
• Judgements that compare many values (is this table sorted, did exactly one thing change) → verify with browser_check or browser_snapshot.

/plugin install jev-browser --marketplace Ying-Kai-Liao/jev-browser

On older Claude Code versions, add the marketplace first:

/plugin marketplace add Ying-Kai-Liao/jev-browser
/plugin install jev-browser@jev-browser

Claude Code then asks for your TypeSafe API key, which goes to its secure credential store. It also offers two optional settings: show the browser window, and a persistent profile directory so that logins survive restarts. The plugin runs the MCP server below, pinned to the same version, and adds a skill that teaches Claude how to write steps and handle each status. The first browser call downloads Chromium for Playwright if it is missing (one time, about 30 s).

Data: see Privacy. Page content goes to the TypeSafe API; passwords and other secrets you pass in values don't.

npx playwright install chromium          # once (the server also does this on first use)
claude mcp add jev-browser -e TYPESAFE_API_KEY=your-key -- npx -y -p jev-browser jev-browser-mcp

Any MCP client works the same way: command npx, args -y -p jev-browser jev-browser-mcp, env TYPESAFE_API_KEY. Add JEV_BROWSER_HEADED=1 to watch it work.

git clone https://github.com/Ying-Kai-Liao/jev-browser && cd jev-browser
npm install
npm run setup                    # downloads Chromium for Playwright
cp .env.example .env             # add TYPESAFE_API_KEY
npm test                         # offline tests (no network, no key)
npm run test:e2e                 # MCP server end to end (network + key)

claude mcp add jev-browser -- node /absolute/path/to/jev-browser/bin/jev-browser-mcp.mjs

From a source checkout the server reads TYPESAFE_API_KEY from the repo's .env.

A page that supports WebMCP registers tools for agents with document.modelContext.registerTool(...) (or navigator.modelContext.provideContext in earlier drafts). An init script records them before the page's scripts run, in every browser — where Chromium has no native modelContext, a minimal one is installed so feature-detecting pages still register. browser_open lists their names as webmcp_tools.

In browser_do, Jev uses them. Each round, the page's tools are offered next to click, type and the rest, with a yes/no "can one of the page's tools do the next step?" that decides between them. The same request fills every tool's arguments: each is one of your values or, for an enum, one of its options (a list, a priority), because Jev never writes text. The page's own execute runs, its result goes into the step's history for the "done?" judgement, and the action shows as webmcp tool add_todo {"text":"values.todo"} => "…". A tool the site does not mark readOnlyHint is guarded like a click: if it looks irreversible the step stops with needs_confirmation.

To call a tool yourself instead, browser_webmcp_tools gives each schema and browser_webmcp_call runs it directly, with no Jev call. Main frame only. Tool descriptions and results come from the site: untrusted, and results are capped at 20k chars. A tool not marked readOnlyHint may have side effects, so the caller should treat it like allow_irreversible.

browser_snapshot is the take-over-and-act surface: numbered elements, plus the text in the viewport. That is the right trade for acting, and it is already small (99–6,900 tokens on every page measured, because elements are capped at 400 and the text at 2,500 chars). It is a poor way to read, though — a Wikipedia article is ~179k chars of text, so a snapshot shows about 1.4% of it and the rest costs a scroll-and-snapshot loop.

browser_read(question) is the understand-the-content surface. Code extracts the whole document's text and cuts it at headings and sections; Jev rates every block against the question in batches of 40; code returns the blocks that help, in document order, inside a character budget. Jev ranks, it never writes — the text you get back is the page's own.

When it pays. Long articles, documentation and reference pages. On the WW2 article, "When and why did Japan surrender?" returned 11.8k of 178.5k chars — 15× less — in 3.4 s over 5 Jev calls, keeping the atomic-bomb and surrender paragraphs.

When it doesn't. Short pages, apps and forms. Below a 8,000-char floor (and below your own max_chars) it returns the page whole and never calls Jev at all, because paying ~1 s to prune a 2k-token page is a loss. Hacker News, a BBC front page and saucedemo all land under the floor.

It is biased to keep: a block Jev cannot rate — or a whole batch whose request failed — is kept, and whatever is left out is counted and marked in place with [… N blocks dropped …]. Nothing is ever silently cut. Use browser_snapshot if you need the page as the caller normally sees it.

node examples/read-demo.mjs                                   # WW2 article, default question
node examples/read-demo.mjs "https://…" "what do you want?"   # needs a key

browser_do statuses:

browser_snapshot options. A snapshot is the one page the calling LLM reads in full, so every option narrows what gets rendered. All of them are pure formatting: nothing is collected differently and none of them costs a Jev call.

A typed password is never printed: input:password values render as [redacted], in a snapshot and in a diff. Measured on a Wikipedia article (1,840 elements), a default snapshot went from 26,919 to 9,926 characters (~6,700 → ~2,500 estimated tokens); filter brings it to ~800 and diff after a scroll to ~50.

Env: JEV_BROWSER_HEADED=1 shows the browser, JEV_BROWSER_PROFILE=/dir keeps a persistent profile (logins survive restarts), JEV_BROWSER_LOG=1 prints per-round decisions to stderr, and JEV_BROWSER_NO_INSTALL=1 turns off the Chromium download on first use. The MCP server reads TYPESAFE_API_KEY from its environment, or from the .env of a source checkout, but never from the .env of the directory it was started in.

// npm install jev-browser && npx playwright install chromium
import { JevBrowser } from "jev-browser";

const b = await JevBrowser.launch({ headed: true });
await b.open("https://www.saucedemo.com/");
await b.do("Log in", { values: { username: "standard_user", password: "secret_sauce" } });
await b.do("Add the Sauce Labs Backpack to the cart");
const p = await b.check("Does the cart badge show 1 item?");   // 0..1
await b.close();

node examples/webmcp-demo.mjs        # headed: 9 plain goals on a planner; Jev picks page tools or clicks, ~25 s
STEP_MS=800 RECORD=demo.mp4 node examples/webmcp-demo.mjs   # same, paced, recorded at 2x (needs ffmpeg)
node examples/x-profile-demo.mjs     # headed, read-only x.com walkthrough with decision highlights

Each action is outlined in red with Jev's choice and scores before it happens (highlight: true, on by default in the MCP server when JEV_BROWSER_HEADED=1). Some sites, x.com included, serve a blank page to headless Chromium: use headed mode there.

node bin/jev-browser.mjs do https://the-internet.herokuapp.com/login "Log in" username=tomsmith 'password=SuperSecretPassword!'
node bin/jev-browser.mjs run examples/flows/todomvc.json --headed

• One observable outcome per step: "Log in", "Open the Pull requests tab", "Mark 'buy milk' completed".
• Every string goes in values with a meaningful key (email, postal_code, file).
• Treat likely_done, needs_confirmation, ambiguous and stuck as your turn: check, snapshot or ask the user, don't just retry.
• After steps with side effects, browser_check what must not have changed.

bin/jev-cost.mjs measures both sides of the trade this project claims to make, on a page you name. Every snapshot, every tool result and every Jev call is measured as it happens — nothing is measured once and multiplied.

npm run cost                                              # the built-in saucedemo + wikipedia cases
npm run cost -- --url https://example.com/ --goal "Log in" --value user=ada --value pw=secret
npm run cost -- --json > run.json                         # machine-readable
npm run cost -- --report run.json                         # re-print the table from a saved run
node bin/jev-cost.mjs --help                              # every flag

Per scenario it reports caller tokens (the text crossing the MCP boundary: every tool result plus the arguments the caller had to write, at chars ÷ 4 — an approximation, and the report says so), Jev tokens and calls (the API's own usage.input_tokens, exact), and wall-clock seconds. The scenarios are run on the same page toward the same goal so they compare: a Playwright-MCP-style baseline that takes a fresh page.ariaSnapshot({ mode: "ai" }) after every action, browser_do, each browser_snapshot option against the default, and browser_read against dumping the whole document.

The net column is caller tokens saved minus Jev tokens spent, one for one, and it goes negative when a scenario loses — on a small page Jev's spend and latency can outweigh the saving, and the tool prints that as a loss. Measured numbers and what they say about the older estimate are in RESULTS.md.

A run needs network and TYPESAFE_API_KEY, and costs API credits. npm test stays offline: the ledger arithmetic is pure and tested in test/cost.test.mjs.

node bench/run.mjs --set all            # base, hard, guard (+ bench/tasks.local.mjs if present)
node bench/run.mjs --only ti-login,drag
node bench/context-cost.mjs bench/results/<run>.json   # older one-sided estimate; see npm run cost

Private sites and credentials go in bench/tasks.local.mjs (git-ignored), exporting LOCAL.

src/session.mjs      JevBrowser: settle, snapshot, decide (1 or 2 stages), resolve, act, do, check, choose
src/page-script.mjs  runs in each frame: elements, labels, state, visible text, metrics, dialogs
src/page-model.mjs   pure helpers: diff between pages, counts, compact rendering
src/webmcp.mjs       init script recording the page's WebMCP tools; result formatting
src/prune.mjs        pure helpers: split page text into blocks, batch, rank, reassemble
src/jev.mjs          System One API client
src/flow.mjs         JSON flow runner
bin/                 CLI and MCP server
.claude-plugin/      Claude Code plugin manifest and the repo's own marketplace
.mcp.json            the plugin's MCP server: npx jev-browser@<version>, key from plugin settings
skills/jev-browser/  skill the plugin adds: how to write steps and handle each status
scripts/             keeps plugin.json and .mcp.json on package.json's version
bin/jev-cost.mjs     two-sided token ledger: caller tokens vs Jev tokens, measured per scenario
bench/               tasks with ground-truth checks, runner, context-cost estimate
test/                offline fixture tests, MCP end-to-end test
NOTES.md             design notes: what works with Jev, what doesn't, and why

CI runs the offline tests on every push and pull request. To publish, bump the version and push the tag; .github/workflows/release.yml tests, publishes to npm with provenance and creates a GitHub release:

npm version patch            # or minor / major: commits and tags vX.Y.Z
git push --follow-tags

npm version also sets the version in .claude-plugin/plugin.json and the pinned jev-browser@X.Y.Z in .mcp.json, so the plugin always launches the release it ships with. CI checks that the three agree and runs claude plugin validate . --strict (npm run validate does both locally). Plugin users get the new version when the marketplace repo updates; the directory listing picks it up from the tracked branch.

Publishing uses npm trusted publishing, set up once with npx npm@latest trust github jev-browser --file release.yml --repo Ying-Kai-Liao/jev-browser --allow-publish.

jev-browser runs locally. It has no server of its own and collects no telemetry. To decide each action it sends the following to the TypeSafe System One API (api.typesafe.ai), using your key:

• the current page's URL, title, visible text and interactive elements, as well as the text blocks browser_read ranks;
• the step's goal, recent action history, and your browser_check and browser_choose questions;
• the values you pass, except those whose key names a secret (password, pw, token, secret, otp, card, cvv, cvc, ssn, api_key, pin). Those are shown by key only. Values typed into password fields are also redacted, and so is a secret value the page shows back (in its text or a WebMCP tool's result).

Playwright types the real values locally. TypeSafe's handling of this data is covered by its own terms. Screenshots and snapshots go only to your MCP client. See PRIVACY.md.

MIT