Item logo image for CUGA Web

CUGA Web

2026latest update
ExtensionWorkflow & Planning
Item media 4 (screenshot) for CUGA Web
Item media 1 (screenshot) for CUGA Web
Item media 2 (screenshot) for CUGA Web
Item media 3 (screenshot) for CUGA Web
Item media 4 (screenshot) for CUGA Web
Item media 1 (screenshot) for CUGA Web
Item media 1 (screenshot) for CUGA Web
Item media 2 (screenshot) for CUGA Web
Item media 3 (screenshot) for CUGA Web
Item media 4 (screenshot) for CUGA Web

Overview

Lightweight browser agent. Controls your tabs with an LLM, supports SKILL.md skills and asks for approval before sensitive actions.

CUGA Web is an AI browser agent in the Chrome side panel. You write a task in plain language. An AI model that you choose does the task in your browser tab: it reads the page, clicks, types, selects options, scrolls and opens pages, one step at a time. You stay in control: the extension asks before sensitive actions, marks the tabs it controls, and never types passwords. You bring the model. Enter the address of any OpenAI-compatible endpoint (for example a LiteLLM proxy or your company's AI gateway), an API key and a model name in Settings. The extension has no server of its own and no account. WHAT YOU CAN DO WITH IT - Fill in long web forms from a short instruction ("File a 49 USD software expense for Ada, dated today"). - Do repeated work in web tools: timesheets, expense reports, tickets, records, settings pages. - Search and read several pages and bring back a short answer. - Teach the agent a task once as a skill, then reuse it with one sentence. - Continue an earlier conversation from the History tab. HOW IT WORKS 1. Click the CUGA Web icon. The side panel opens. 2. Type a task in the Chat tab and press Send. 3. The agent works in the tab you are looking at. After every action it reads the page again and decides the next step. 4. Before a sensitive action, such as Submit or Delete, a card in the chat asks for your approval. 5. When the task is complete, or cannot be done, the agent gives a short summary. The agent sees a page as a numbered list of its interactive elements (buttons, links, inputs, options) plus the visible text. It can also see a screenshot of the visible area if your model supports images. It then calls one of these actions: read the page, click, type, select an option, press a key, scroll, open a URL, go back, open or switch tabs, take a screenshot, wait, load a skill, save a skill, ask you a question, or finish. APPROVAL BEFORE SENSITIVE ACTIONS The approval check runs in the extension code. It is not an instruction to the model, so the model cannot skip it. Three modes (Settings > Approval): - Sensitive (default): asks before a click on an element whose text, label, value or title contains a sensitive word; before a click on a submit button inside a form; before Enter in a form field; and before typing into a password field. - Always: asks before every click, typing, selection and Enter key press. - Never: asks for nothing. Use it only on pages you do not care about. The default word list is: save, submit, delete, remove, send, pay, buy, purchase, confirm, publish, post, order, sign. You can edit the list. Matching is by whole word and ignores upper and lower case. The approval card shows what the agent wants to do, which element, which site and why it asked. You can press Approve, Deny, or "Always allow here" (adds the site to a list in Settings; you can remove it there). After Deny, the agent is told not to try the action again or work around it. CLEAR MARKING OF CONTROLLED TABS You can always see which tabs the agent controls: - The tabs are put in a Chrome tab group named CUGA. - The toolbar icon shows the badge AI on those tabs. - The page shows a teal border with the text "CUGA is controlling this tab". The border is hidden while a screenshot is taken, so it does not appear in the image. - Tabs that the page opens during a task join the group. Everything is released when the task ends. The extension reads only the tabs it controls, and only while a task runs. SKILLS: TEACH A TASK ONCE A skill is a short text file named SKILL.md. It has a name, a one-sentence description of when to use it, and numbered steps. At the start of a task the agent sees only the names and descriptions of your skills. When a task matches, it loads the full text and follows it. This keeps requests small. Ways to add a skill: - Skills tab > Add skill: opens an editor with a template. The editor checks the text while you type. - Import a SKILL.md file, or drop the file on the panel. - Ask in the chat: "Save this task as a skill" or "Update the expense-report skill: always ask for the date". - Press "Save as skill" under a finished task. Creating and updating skills by chat uses the built-in skill named skill-manager. It works like this: the agent asks you 2 to 4 questions in one card (where the skill applies, which values stay the same every time, what to do when a value is missing, which actions need your OK first). Then it writes the skill and shows the full text with the buttons Save, Edit first and Cancel. Nothing is stored before you press Save. This card appears in every approval mode. The agent creates or changes a skill only when you ask. It may suggest one after a long task, at most once in a conversation. You can turn skills on and off, edit them, export them as files and delete them (the delete button asks for a second click). The built-in skill can be viewed and turned off, but not edited or deleted. The skill format is the same as in CUGA agent projects, so a skill file can be shared between them. QUESTIONS IN THE CHAT When the agent needs information, it does not guess. It asks in the chat with a card: - Questions with fixed answers show options to click: one choice or several, plus "Other" with a text field. - Open questions show a text box. - With more than one question, the card shows progress and has Next, Back and Submit. - You can also type your answer in the message box at the bottom. LOGIN PAGES The agent does not type or guess credentials. When a page asks to sign in (login form, single sign-on, verification code, "session expired"), it stops and asks you to sign in on the page yourself. After you press "I signed in, continue", it reads the page again and continues from where it stopped. The values of password fields are never read. HISTORY The History tab lists your last 30 conversations with date, number of tasks and result. Click one to open it and continue. The agent reads the page again when you continue. Conversations are stored only in your Chrome profile, without screenshots. Delete one conversation or clear all (both ask for a second click). You can turn history off in Settings. SETTINGS - Base URL, API key and model name of your OpenAI-compatible endpoint. - Test connection: runs three checks and shows pass or fail for each: plain chat, tool calling (function calling), and image input. If image input fails, one button turns screenshots off, and the agent then works from page text only. - Screenshots: after every action, only when the model asks, or never. - Approval mode, the sensitive word list and the list of sites where you are never asked. - Maximum steps per task (default 30, up to 100). - Save conversations for the History tab (on or off). A model that supports tool calling is required. Image input is optional but improves results on pages with unusual layouts. EXPORT LOG FOR SUPPORT AND DEBUGGING The Export log button in the Chat tab downloads a JSON file with the events of the current session: your task, each model request and response, each action and its result, approvals and errors. Your API key is replaced by the word "set" and screenshots are left out. The file contains task text and page text, so check it before you share it. The log is kept in memory and is lost when the side panel closes. YOUR DATA AND PRIVACY - The publisher of this extension receives no data. There is no server, no analytics, no tracking and no ads. - During a task, the following is sent to the model endpoint that you entered in Settings, and to no other place: your task text and answers, the URL, title, visible text and list of interactive elements of the controlled tabs, screenshots of those tabs if screenshots are on, and the names and descriptions of your skills (the full text of a skill when the agent loads it). - Settings, skills and history are stored in your Chrome profile with chrome.storage.local. They are not synchronized. The API key is stored unencrypted in the profile. - Page text can include personal data that the page shows. Do not use CUGA Web on pages you do not want to send to your model endpoint. - The provider of your endpoint processes the data under its own terms. - Removing the extension deletes everything it stored. Privacy policy: https://cuga.dev/cuga-web/privacy.html WHY THE EXTENSION ASKS FOR ITS PERMISSIONS - Side panel: the whole interface and the agent loop run in the Chrome side panel. - Storage: saves your settings, skills and history in your browser. - Tabs: reads the URL, title and loading state of the controlled tab, opens tabs when a task needs it, and takes screenshots of the controlled tab. - Tab groups: creates the CUGA tab group so you can see which tabs are controlled. - Scripting and access to all sites: runs the extension's own bundled code in the controlled tab to read the page and perform clicks, typing, selections and scrolling. You can ask the agent to work on any site, and your model endpoint can be on any host (for example a local proxy), so these cannot be limited to a fixed list. The extension acts on a page only while your task runs in that tab. - No remote code: all code is inside the extension package. The model returns action requests as data (an action name and its arguments), and the extension maps them to its own fixed functions. Model output is never run as code. SAFETY NOTES - Web pages can contain text that tries to give instructions to an AI agent. The agent is told to treat page content as data and never to follow instructions found on pages, and sensitive actions need your approval. This reduces the risk of such attacks but does not remove it. Keep the approval mode on Sensitive or Always for important accounts. - Review the approval card before you press Approve. The agent can make mistakes. - Do not use the extension for actions you are not allowed to automate on a site. Follow the terms of the sites you use. REQUIREMENTS - Chrome 116 or newer. - An OpenAI-compatible chat completions endpoint with a model that supports tool calling. - An API key for that endpoint, if it needs one. LIMITS - Chrome's own pages (chrome://) and the Chrome Web Store cannot be controlled. If the active tab is such a page, CUGA opens a new blank tab. - Clicks and typing are sent as page events. A few sites ignore such events. Masked input fields (for example time fields in some business apps) get a second typing attempt with one key event per character; the action result tells the agent what the field shows afterward. - Elements inside closed shadow DOM are not read. Open shadow DOM and iframes (including cross-origin iframes) are read. - Elements below or above the visible area are listed when scrolling reaches them. Elements that scrolling cannot reach, such as options of a long dropdown list, are listed as offscreen and can be clicked directly. - Screenshots show the visible area of the active tab. CUGA activates the controlled tab before a screenshot. - Closing the side panel stops the running task. - The quality of the results depends on the model you choose. QUESTIONS AND ANSWERS Does it work without my own model? No. You need an OpenAI-compatible endpoint and a model. The extension does not include a model. Does it work with a local model? Yes, if the endpoint is OpenAI-compatible and the model supports tool calling, for example a LiteLLM proxy at http://localhost:4000/v1. Which model name do I enter? The name or alias that your endpoint uses. With LiteLLM, this is the model_name from your config file. Can it see my other tabs? No. It reads only the tabs in the CUGA group, which it controls for the current task. Can it log in for me? No. It asks you to sign in and continues afterward. Where are my skills stored? In your Chrome profile. Use Export on a skill to save it as a file. How do I report a problem? Use Export log in the Chat tab, check the file, and send it with a short description of what you saw.

Details

  • Version
    0.2.0
  • Updated
    November 19, 2026
  • Offered by
    sami.marreed.16@gmail.com
  • Size
    91.99KiB
  • Languages
    English
  • Developer
    Email
    sami.marreed@ibm.com
  • Non-trader
    This developer has not identified itself as a trader. For consumers in the European Union, please note that consumer rights do not apply to contracts between you and this developer.

Privacy

Manage extensions and learn how they're being used in your organization

CUGA Web has disclosed the following information regarding the collection and usage of your data. More detailed information can be found in the developer's privacy policy.

CUGA Web handles the following:

Personally identifiable information
Web history
Website content

This developer declares that your data is

  • Not being sold to third parties, outside of the approved use cases
  • Not being used or transferred for purposes that are unrelated to the item's core functionality
  • Not being used or transferred to determine creditworthiness or for lending purposes

Support

For help with questions, suggestions, or problems, visit the developer's support site

Search
Clear search
Close search
Google apps
Main menu