Releases·Tool Release
jev-browser: TypeScript browser automation where Claude plans and Typesafe System One's Jev picks actions
Two related open-source projects build browser agents around Jev (Typesafe System One), a choice model that picks actions from indexed page candidates rather than generating them. fastbrowse reports 41/42 task passes at $0.0057 median cost per task, 71x cheaper than Browser Use hosted, while jev-browser reports 40/42 with zero false 'done' claims. Both remain experimental and rely on the TypeSafe API.
116 engagement2 sourcesWed, Sep 16, 2026, 03:13 PM
AI Overview
- jev-browser (GitHub, 2026-09-16) splits browser automation: the calling LLM plans outcomes one step at a time while Jev decides each action; unofficial, not affiliated with TypeSafe
- Ships as a library, CLI, and MCP server; each step loops through network/DOM settle, page description, a ~300 ms Jev request, then Playwright execution
- Each Jev call answers multiple questions at once: which element, which action, which value, and whether the step is done, blocked, erroring, or about to act irreversibly
- Jev returns only probability distributions (yes/no 'noul', 'choice', 'score'), never free text; the caller supplies candidates and code converts low confidence into actionable status
- The LLM never reads page snapshots unless it takes over; it receives { status, url, actions[], done_score } per step
- On 42 tasks across 16 categories on live sites, the latest run scored 40/42 correct with zero false 'done' claims, up from 38/41 twice before fixes; misses were counting and verifying a sort
- fastbrowse (GitHub, 2026-09-17, pre-alpha, Python 3.13/3.14, MIT) cites a verbatim page quote for every claim and can't click candidates never on the page
- On 14 answer tasks x3 passes uncapped, fastbrowse passed 41/42 at $0.0057 median/task and 21.0s median vs Browser Use's 42/42, $0.41, 24.8s — 71x cheaper, 1.2x faster
Related
Products·launched
Hunch: open-source browser agent using ~150ms model decisions instead of LLM calls per step
Hunch is an open-source, zero-dependency browser agent that replaces per-step LLM reasoning calls with a small model that returns typed action probabilities in ~150ms. It reads accessibility snapshots from agent-browser, picks the operation and target element, and verifies outcomes in code. This addresses the 30-60s latency of typical LLM-driven login and form flows for developers building browser automation.
Hunch is an open-source, zero-dependency browser agent that replaces per-step LLM reasoning calls with a small model that returns typed action probabilities in ~150ms. It reads accessibility snapshots from agent-browser, picks the operation and target element, and verifies outcomes in code. This addresses the 30-60s latency of typical LLM-driven login and form flows for developers building browser automation.
18 engagement1 sourceredditMon, Sep 21, 2026, 11:51 PM
Products·building
IronBee Express - fast browser agent that executes tasks and verifies results
IronBee Express is a browser agent: you describe a task in plain English and it carries it out in a real browser, then reviews whether it actually succeeded. It uses LLMs (via Jev) to drive actions and to verify outcomes by checking page state against API responses, flagging runs as failed on mismatch. It targets people who need reliable, fast automation of web tasks like sign-ins and checkouts rather than demos that only look successful.
IronBee Express is a browser agent: you describe a task in plain English and it carries it out in a real browser, then reviews whether it actually succeeded. It uses LLMs (via Jev) to drive actions and to verify outcomes by checking page state against API responses, flagging runs as failed on mismatch. It targets people who need reliable, fast automation of web tasks like sign-ins and checkouts rather than demos that only look successful.
0 engagement1 sourceredditThu, Oct 1, 2026, 03:29 PM
Products·building
Local web browsing agent using only the Clef-Flash 9B decision model
A system that performs web browsing tasks using only a small local decision model, Clef-Flash 9B, which cannot generate text, so page text is extracted at runtime to inform decisions. Demonstrated on tasks like finding release downloads and usage-limit reset times, completing them in 13-17 seconds. It targets people who want private, lightweight browser automation without a large LLM.
A system that performs web browsing tasks using only a small local decision model, Clef-Flash 9B, which cannot generate text, so page text is extracted at runtime to inform decisions. Demonstrated on tasks like finding release downloads and usage-limit reset times, completing them in 13-17 seconds. It targets people who want private, lightweight browser automation without a large LLM.
1 engagement1 sourceredditSun, Oct 11, 2026, 04:00 PM
Social·Community
Jev benchmark defines truth as GPT-6 Astra and Claude Fable 5.1 agreement, scores Opus 5 at 73.1%
A Reddit post critiques TypeSafe's Jev benchmark, arguing its reference labels come from an average of GPT-6 Astra and Claude Fable 5.1 responses, making scores measures of agreement with Claude and GPT rather than accuracy. The post also disputes the advertised 0% hallucination rate, which is defined only as schema conformity, and notes TypeSafe's own jaggedness page for jev-1.13 lists nine failure modes. It contrasts this with the author's open-source validator toolchain, whose SHACL and OWL-RL results are machine-checked in Lean.
A Reddit post critiques TypeSafe's Jev benchmark, arguing its reference labels come from an average of GPT-6 Astra and Claude Fable 5.1 responses, making scores measures of agreement with Claude and GPT rather than accuracy. The post also disputes the advertised 0% hallucination rate, which is defined only as schema conformity, and notes TypeSafe's own jaggedness page for jev-1.13 lists nine failure modes. It contrasts this with the author's open-source validator toolchain, whose SHACL and OWL-RL results are machine-checked in Lean.
1 engagement1 sourceredditSun, Sep 20, 2026, 10:25 AM
Social·Community
Web browsing was implemented using only the Clef-Flash 9B decision model, running locally.
A video demonstrates web browsing running entirely on a locally hosted Clef-Flash 9B decision model, completing three tasks in 11–17 seconds each. Since the model cannot generate text, the system extracts text from web pages at runtime for decisions, with commands entered manually in the demo though an LLM could automate them. Results were strong on simple sites but degraded on complex UIs.
A video demonstrates web browsing running entirely on a locally hosted Clef-Flash 9B decision model, completing three tasks in 11–17 seconds each. Since the model cannot generate text, the system extracts text from web pages at runtime for decisions, with commands entered manually in the demo though an LLM could automate them. Results were strong on simple sites but degraded on complex UIs.
9 engagement1 sourceredditSun, Oct 11, 2026, 04:17 PM