Koretex Browser Agent

About

Koretex Browser Agent is an AI agent that lives in Chrome's side panel. Type a task in plain language — "find 5 second-degree decision makers in Melbourne on LinkedIn and copy their details into a Google Sheet" — and it drives your browser to deliver it: navigating, clicking, typing, reading, and checking its own work from screenshots at every step.

HOW IT WORKS

• Judge, don't assume. After every action the agent looks at an actual screenshot and rules whether the action worked — success is something it must see, never something it predicts.
• Safety lives in code, not prompts. Anything irreversible (posting, deleting, purchasing) is guarded by the runtime itself: one attempt, no blind retries, honest stops.
• Knowledge is taught, not hardcoded. The agent carries "skills" — small playbooks of site knowledge — and you can teach it new ones by demonstrating a task while it watches. Skills are editable and shareable as plain JSON.

You watch a live trace of everything it does: each action, its judgment of the outcome, which playbooks are in force, and exactly what each decision cost. Nothing is hidden. A typical task costs $0.002–$0.04 in model calls and takes seconds to a few minutes.

If a run gets stuck, it doesn't flail: stuck-detection triggers a strategic review that diagnoses the root cause and charts a different route. If it still can't finish, it stops honestly and reports what it did gather.

WHAT YOU NEED

Bring your own model access: an API key for cloud models, and optionally a local Ollama install for on-device click grounding and page reading. A fully local mode (no API key, reduced capability) is also available.

PRIVACY — READ THIS

• In the default mode, screenshots of your active tab are sent to a cloud model at every step, including whatever is visible in your logged-in session. All cloud calls request no-data-retention routing.
• Local models (via Ollama) handle click grounding and bulk page reading, so full page text does not leave your machine in the default mode.
• The agent never logs in or handles passwords — if a login wall appears, it stops and tells you.
• Optional cloud-only mode adds a PII guard (identifiers are replaced with tokens before text leaves the machine) and asks before working on banking, health, or government sites.
• Full policy: https://github.com/koretex-ai/koretex-browser-agent/blob/main/PRIVACY.md

WHY THESE PERMISSIONS

• Access to all sites: the agent operates on whatever site your task requires.
• Debugger: trusted keyboard input for canvas editors (Google Docs/Sheets) and handling native browser dialogs — synthetic events are ignored by exactly the apps that matter. Chrome shows a debugging banner while the agent runs.
• Tabs, scripting, storage, alarms: driving the active tab, saving your skills and chat history, and running your scheduled tasks.

Open source (Apache 2.0): https://github.com/koretex-ai/koretex-browser-agent
Community: https://discord.gg/4xtjpKf5p