Do I need my own model key and a specific browser setup?

Yes. You supply your own Gateway key in side-panel Settings, run a Chromium-based browser with side panels and the debugger API, and install Bun to build. Development was tested in Helium on macOS.

How is this different from scripting clicks with Stagehand?

Ulka keeps local browser ownership, approval gates, evidence capture and final verification, but adopts observe, act, extract and cached replay primitives. Successful clicks can replay only when URL, goal, page state and target meaning match.

What actually leaves my machine when a task runs?

Page text, element labels, field values, task history and messages may go to Gateway and its providers. The key, chats and diagnostics stay in chrome.storage.local, which is not an encrypted vault.

Can it keep working while I browse another tab?

Yes, if you pick Tab behavior Background before starting. It keeps its task target and switches targets without activating tabs, but hidden-tab pauses are possible and closing the task tab stops work.

How do I build and load the extension?

Install Bun 1.4.2, run bun install --frozen-lockfile and bun run build, then load apps/extension/dist unpacked from your browser's extensions page and paste your Gateway key in the side panel's settings.

Which models does it actually call?

Planning and verification use deepseek/deepseek-v4.1-flash with reasoning disabled, while action selection uses typesafe-ai/jev. Model availability and cost depend on your Vercel AI Gateway key.

Are there limits on how long a task runs?

Plans with at least three milestones get a 48-tool-call ceiling and ordinary tasks 24, with execution bounded by shared 30-action and 60-runner-model-step allowances.

CommunityJev Demos & Experiments

Ulka

Local browser-agent extension where FX plans tasks and Jev picks constrained actions from observed elements, with model calls routed through Vercel AI Gateway.

See highlights

Open github.com

Stars: 22
Language: TypeScript
License: MIT
Published by: Community
Author: razaanstha
Last verified: Oct 2026

razaanstha/ulka

User repository on GitHub

View on GitHub

Experimental browser agent powered by FX, Jev, and Vercel AI Gateway. Bring your own API key to read pages and automate browser tasks.

Forks: 4
Last pushed: Updated Oct 2026

Repo stats from the GitHub API, cached Oct 2026.

About this project

FX orchestrates browser tasks and tool calls. Jev (typesafe-ai/jev) selects constrained browser actions from observed page elements. Vercel AI Gateway routes AI requests using the user's API key. Ulka browser runtime reads Chromium’s accessibility tree for roles, names, states, and UI hierarchy, validates targets, requests approvals, executes actions, and records evidence.

Quoted from the project's README. Read the full README.

Tagscommunity jevbest catalog-import

Works forJev for Classification Jev for Tool calling

Used inJev for Software & DevTools

Languages: TypeScript

Highlights

Reads the Chromium accessibility tree for roles and hierarchy, validates targets and gates actions behind approvals with evidence capture.
Stagehand-style observe, act, extract and cached replay with milestone plans and bounded tool-call and action allowances.
Runs locally with no app backend; load apps/extension/dist unpacked and save the Gateway key in side-panel Settings.
Supports background tab mode, WebMCP tool preference, and personalization text sent with new tasks.

Watch out

Early preview with expected bugs; supervise signed-in sites, approvals are heuristic and page text is sent to model providers. Copy reviewed against the repository README on 2026-10-04, not from running it.

Questions about this entry

Short answers drawn from the linked project or page.

More Jev resources like this

jev-browser-skill

Standalone CLI skill that drives isolated Playwright Chromium with Jev choosing scoped in-page actions, defaulting to Vercel AI Gateway with TypeSafe as an option.