Show HN: Jev-browse – browser sub-tasks for coding agents at ~1/3 the cost

Fast, typed browser sub-tasks for coding agents. jev-browse adds a few helpers to browser-harness so that Claude Code or Codex can hand off a whole website sub-task in one call: "search Lisbon, set these two filters, open Casa Flora". Each step is decided by TypeSafe Jev, a small model that answers typed questions in a few hundred milliseconds. The agent then gets the outcome, or a typed reason why it stopped.

r = fast_run("https://en.wikipedia.org/wiki/Main_Page",
             "Open the Wikipedia article on Gödel's incompleteness theorems", run_id="demo-1")
print(r.status, r.reason, r.url)      # claimed_done None https://en.wikipedia.org/wiki/G%C3%B6del%27s_...
jev_close(r.target_id)

Why. When a coding agent drives a browser itself, each click costs a model turn that re-reads its whole context. A five-field form becomes fifteen turns. fast_run does the clicking and typing in one call, with one Jev request per decision, and hands back to the agent only when it must: a value it doesn't know, a risky click, a frame or upload it can't handle, or anything sensitive. In the author's runs that made form and navigation tasks 1.6–3.0× faster and 2.1–5.3× cheaper than the agent driving browser-harness alone, with the same pass rate (details).

It is a port of browser-use/jev-ultrafast (MIT), turned into helpers any agent can call from a browser-harness script.

You need Chrome, browser-harness connected to it, Python 3.11+, and a TypeSafe API key from console.typesafe.ai/keys.

Paste this into Claude Code or Codex:

Install jev-browse from https://github.com/danielnc/jev-browse by following its install.md: clone it, run its
installer, store my TypeSafe API key in the browser-harness agent-workspace .env (ask me for it; never print it),
and run `python3 -m jev_browse doctor` until it passes. Ask me which text backend I want (default: my Claude
subscription if the claude CLI is installed) and whether to add the jev-browse pointer to my global agent
instructions.
git clone https://github.com/danielnc/jev-browse ~/jev-browse && cd ~/jev-browse
python3 -m jev_browse install          # adds the helpers to browser-harness and links the skill
ENV=~/.config/browser-harness/agent-workspace/.env
printf 'TYPESAFE_API_KEY=%s\n' '<your key>' >> "$ENV" && chmod 600 "$ENV"
python3 -m jev_browse doctor           # checks everything and prints what is active
browser-harness <<'PY'
r = fast_run("https://en.wikipedia.org/wiki/Main_Page", "Open the Wikipedia article about the Eiffel Tower",
             run_id="hello-1")
print(r.status, r.url)
jev_close(r.target_id)
PY

Then add the global pointer to your agent's instructions. Without it, agents rarely think to use jev-browse on their own.

 your agent (Claude Code / Codex)
   │  writes one browser-harness script:  r = fast_run(url, goal, values={...})
   ▼
 browser-harness  ──CDP──►  Chrome: a new background tab owned by jev-browse
   │                           │
   │   ┌───────────────────────┘
   │   ▼
   │  loop:  snapshot the page (snapshot.js: visible text, controls, fields; sensitive values never read)
   │         → one TypeSafe Jev request: next operation? which target? which value?  ──►  api.typesafe.ai
   │         → value missing from the goal?  ask the text backend once (optional)   ──►  Claude / Codex /
   │         → safety gates: commit verbs, sensitive fields, frames, uploads, hosts        Ollama / any
   │         → click / type / select / scroll via CDP                                     OpenAI-compatible
   │  until DONE, BLOCKED, or a typed hand-back (confirm_required, in_frame, text_value_unavailable, …)
   ▼
 RunResult(status, reason, url, evidence, trace)  +  a JEV_BROWSE_RESULT={...} line to verify in the same script

• One decision = one small Jev request (typically a few hundred milliseconds). No screenshots and no large model in the loop.
• Values come from yourvalues= , from the goal itself (Jev picks among candidates taken from the goal), or, only after a miss, from a text backend. A grounding gate checks the backend's answer before anything is typed.
• It never guesses past its limits. Too many options, frames, shadow DOM, canvas, uploads, missing values, sensitive fields, or a click that would send, pay, delete, or book all hand back with aReason . The tab stays open for the agent to finish or resume.

More: docs/architecture.md.

When fast_run hands back, the agent continues with the harness on the same tab, or resumes after supplying what was missing: fast_run(None, goal, target_id=r.target_id, values={...}).

Zero config beyond TYPESAFE_API_KEY. Everything else is an environment variable (the harness .env counts) or an entry in ~/.config/jev-browse/config.toml. The environment wins. The settings you are most likely to change:

Privacy note on text.fallback. With a local ( ollama) or OpenAI-compatible backend, the default text.fallback = "auto" means: if that backend fails its known-answer check, is unreachable, or returns invalid output twice, the claude CLI answers instead (when it is installed), so the goal, field labels, and page excerpt go to Anthropic. If you chose a local model to keep page text on your machine, set text.fallback = "none" (JEV_BROWSE_TEXT_FALLBACK=none): the miss then hands back to your agent. doctor prints which fallback is active.

All settings, and common setups (privacy mode, local model, OpenRouter/Groq/Cerebras/Gemini): docs/configuration.md. python3 -m jev_browse config shows what is active and where each value came from.

A text backend is asked only when a field's value is implied but not stated ("the capital of France" → Paris).

Personal fields (name, email, phone, address) and sensitive fields are never sent to a text backend. Honest numbers and set-up notes: docs/backends.md.

• What leaves your machine. On every decision, visible page text, element labels, non-personal field values, URL, and title go toTypeSafe (api.typesafe.ai ). After a miss, the goal, field labels, and up to 2,000 characters of page text go to yourtext backend , or to its fallback when a local backend fails (seetext.fallback above). Use it only on pages you are comfortable sending there. Usetext_backend="none" (ortext.backend = "none" ) for sensitive sites.
• Sensitive fields (passwords, one-time codes, card numbers, CVV, IBAN, national IDs) are never typed, never read out of the page, and never sent. Personal fields are reported only as filled or empty.
• Commit gate. Clicks that send, pay, delete, book, or confirm need the goal's explicit authorisation, or yourconfirm=[...] in alater script. Otherwise the run hands backconfirm_required . This is a heuristic (verbs plus structural signals such as a confirmation inside a dialog): it lowers the risk of an unwanted click in your signed-in browser. It does not remove it.
• Owned tabs. jev-browse works only in background tabs it created (or that yournew_tab() created and you adopted). It never touches, focuses, or closes your other tabs.
• Hosts.safety.allowed_hosts restricts every page jev-browse opens, adopts, or observes.
• No telemetry in jev-browse. browser-harness has its own telemetry, which sends script text and helper-call arguments to PostHog when enabled. The installer anddoctor warn if it is on. Opt out withbrowser-harness telemetry disable .
• Local data. Run files, traces, and screenshots can contain page text and typed values. They stay in the harness tmp dir and other gitignored paths, andmake clean-traces removes them.

A skill that never triggers delivers nothing. With only the skill installed, the calling agent used jev-browse in 0 of 5 unprompted tasks in the author's runs. After a short pointer in the global instructions it used it in 2 of 5, and the explicit fast_run call shape saved a turn per task. Paste this into ~/.claude/CLAUDE.md, ~/.codex/AGENTS.md, or your project's agent file:

## Browser tasks: jev-browse fast path
For a multi-step website sub-task (search, fill, filter, open a result) where you only need the outcome, call
jev-browse from a browser-harness script. You don't need to load its skill first:

    r = fast_run(url, goal, values={...known field values...}, run_id="<unique>")
    print(r.status, r.reason, r.detail, r.target_id)

- `claimed_done` is not proof: check the printed JEV_BROWSE_RESULT line, or `js("...", target_id=r.target_id)`,
  in the same script, then `jev_close(r.target_id)`.
- On a hand-back (`r.reason`), load the jev-browse skill for what to do next. Resume on the same tab with
  `fast_run(None, <same goal>, target_id=r.target_id, values={...})`; never re-run from the URL.
- Never act on `confirm_required` in the same script: decide first (ask me if my request does not clearly cover
  it), then resume with `confirm=[...]`.
- Use browser-harness directly for visual judgement, frames, uploads, and sensitive pages. Page text goes to
  TypeSafe; pass `text_backend="none"` on sensitive sites.
- Close only tab ids that jev-browse returned to you.

The same snippet is in docs/global-pointer.md.

• No iframes, shadow DOM, canvas or visual understanding, file uploads, or pop-up tabs: these hand back.
• Pages with very many options or very large state hand back (too_many_options ,state_too_large ).
• Decisions are text-only. A page whose meaning is visual will confuse it, and it will usually say so (visual_only ,low_confidence ).
• Every decision is a metered TypeSafe request: a few tenths of a cent per task at the time of writing.
• The commit gate and the grounding gate are heuristics. Review what a run did (r.trace ) on anything that matters.
• The Codex and OpenAI-compatible backends are unmeasured.
• Tested with browser-harness 0.1.13 on macOS with Chrome. Other platforms should work but have not been benchmarked.

jev-browse was compared with the same coding agent driving browser-harness directly, on public-site navigation, form, and filter tasks, with every outcome verified by code. In the author's runs it was 1.6–3.0× faster and 2.1–5.3× cheaper, with the same pass rate (N = 3 each), mostly because the agent needs far fewer turns. Your numbers will differ. Summary: docs/benchmark.md. Reproduce it, or evaluate your own text backend, with bench/ (docs/benchmarking.md).

• install.md : step-by-step install, written for an agent to follow
• docs/architecture.md : the decision loop, modules, and safety model
• docs/configuration.md : every setting
• docs/backends.md : text backends and their trade-offs
• docs/benchmarking.md : running the benchmark andtext_eval
• skill/SKILL.md andskill/reference.md : what the agent reads

See CONTRIBUTING.md. Development is offline by default: uv sync && make check runs the linter and more than 400 tests with TypeSafe, the CLIs, and Chrome mocked.

MIT, see LICENSE. jev-browse ports code from browser-use/jev-ultrafast (MIT, Copyright (c) 2026 Browser Use); see NOTICE. Ported files carry a header comment naming the upstream file.