Web Draw by Olib AI

Let an AI agent read and drive the page you are looking at, as text, without screenshots.

Web Draw lets an AI assistant read and operate the browser tab you already have open, using text instead of screenshots. It works with LLM-based assistants and now offers optional support for System One controllers and TypeSafe's Jev model. The extension turns the visible page into a compact text view: buttons, links, fields and menus with short action handles, required or invalid fields, disabled controls, dialog text and responses to clicks. It uses your existing browser sessions, with no separate login or need to copy session credentials. HOW IT WORKS Web Draw has two parts. This extension reads the page and performs browser actions. A small local server connects it to your AI client over the Model Context Protocol (MCP). After installing, add the server to your MCP client configuration: "web-draw": { "command": "npx", "args": ["-y", "@olib-ai/web-draw-mcp"] } The toolbar badge shows ON once the two are connected. You need Node.js 20 or newer and an MCP-compatible AI client. WHAT IT IS GOOD AT - Filling and submitting forms, including long multi-step ones - Reading dense pages such as search results, dashboards and data tables - Handling dialogs, cookie walls, overflow menus and controls inside iframes - Reporting when a page rejects a submission and which fields need attention - Giving agents actionable controls and page state without screenshot interpretation OPTIONAL SYSTEM ONE / JEV SUPPORT Version 1.1.0 adds structured page observations and explicit action choices for decision models. The existing five-tool LLM workflow remains the default. Enable the additional tools by setting WEB_DRAW_SYSTEM_ONE=1 in the MCP server environment. Use version 1.1.0 or later of both the server and extension, then restart your MCP client. - browser_observe returns structured page state and available action candidates. - browser_decide asks TypeSafe Jev to propose a candidate for a goal, with probabilities and confidence. It does not execute the action. - browser_execute_choice executes a selected candidate and returns the outcome and an updated observation. The built-in Jev adapter requires your own TYPESAFE_API_KEY in the server environment. Structured observations and candidate execution do not require a TypeSafe key and can be used by another controller. Jev chooses among supplied actions. Your assistant or controller provides any text to type and decides when to execute a proposal. Web Draw checks that the selected target still matches the observation and rejects expired or already-used snapshots. Controllers should verify the resulting page before declaring a task complete. PRIVACY Browser reading and execution happen on your machine. The extension communicates with your configured local bridge. Page content is passed to your AI client, which may send it to its model provider according to your client settings. If you enable and call browser_decide, the local server sends the goal, page observations and candidate descriptions to TypeSafe for evaluation. The TypeSafe API key stays in the server environment. The other Web Draw tools do not call a model provider. Nothing is sent to Olib AI, and this extension has no analytics or tracking. WHY THE PERMISSIONS - Access to sites and tabs: to read the page your assistant requests and act on its controls. - Scripting: to load the reader into frames that do not have it yet, including pages open before installation. - Debugger: a fallback for the standard browser tools when a page ignores synthetic input. Your browser shows its standard notification while this is active. - Web navigation: to find frames inside a page. - Storage and alarms: to keep connection settings and maintain the local connection. LOOKING FOR HEADLESS AUTOMATION? For large-scale headless automation, parallel browser contexts or antidetect browsing, see Owl Browser: https://www.owlbrowser.net