Codex Can Now Work Inside Users’ Live Chrome Sessions

Dominik Kundel presents the Chrome extension as a way for Codex to work where much of a user’s work already exists: inside Chrome, with the same profile, session, cookies, tabs, and logged-in apps. The extension works with the Codex app on both Windows and macOS.

The product distinction is not that Codex lacked browser access before. Kundel says the in-app browser remains useful for local development tasks and for giving Codex detailed feedback through annotations. The extension is aimed at a different problem: work that depends on the user’s existing Chrome state or on capabilities available only in a full web app.

The installation view in the Codex app describes the extension as a way for Codex to use “open tabs or websites you’re already signed into.” The visible product copy says it can “navigate, inspect, click, type, and take screenshots while working.”

Kundel does not frame Chrome control as the first choice for every task. If there is a plugin for the app he needs, he says he normally starts there. A Codex plugin view lists featured plugins including Computer Use, Browser Use, Spreadsheets, Presentations, GitHub, Slack, Notion, Linear, Statsig, Gmail, Google Calendar, Google Drive, Teams, and SharePoint.

The reason is speed and structure. Connectors are “structured and fast,” Kundel says, because Codex does not have to click through a user interface just to read a document, check a message, or create a file. The Chrome extension is for the cases around that model: when there is no plugin, when a plugin exists but the needed operation is available only in the full web app, or when the relevant context is the already logged-in browser session.

The operating pattern is straightforward: use structured plugins where they fit, then use Chrome when the task requires browser-native state or interaction with a live web application.

A central claim for the extension is that Codex can work across Chrome without taking over the user’s whole browser. Kundel contrasts it with regular computer use: Codex “won’t take over my whole browser.” Instead, it creates its own Chrome tab group, works from there, opens multiple tabs, scrolls through pages, finds content, and reasons as it goes.

The research example shows what that means in practice. Codex is asked to review the last week of Codex-related posts on community.openai.com, summarize user feedback, assess sentiment around recent launches, identify key user issues, and output the findings into a spreadsheet.

The browser view shows OpenAI Developer Community search results for “codex” over the period after 2026-04-28 and before 2026-05-06, with “50+ results.” Codex’s working log shows it inspecting the page and adapting when one route fails: a visible note says a JSON search route was blocked in Chrome by the page client, so Codex switched to the rendered forum search and topics page.

The completed task message says Codex worked for 9 minutes and 51 seconds and created and verified a spreadsheet covering Apr. 28–May 5, 2026. It says the spreadsheet includes 67 forum result rows and 34 high-signal topic deep reads, separated into summary, themes, key issues, launch sentiment, source posts, and deep-thread evidence.

For the user, the point is that this work can run in Codex’s own Chrome tab group, in the background, while the user keeps working in other tabs.

The expense-report example combines Chrome control with a plugin and local files. Kundel says Codex can check email through a plugin, search for messages related to a recent trip, extract the relevant information, and then fill out expense forms.

The prompt asks Codex to check Gmail for food-related expenses from a recent Portland trip and add them to expenses in Navan, with related receipts stored in ~/Desktop/receipts. The reasoning log says Codex found two local receipt PDFs, hey-love.pdf and bun-mee.pdf, and is checking Gmail for trip context and receipt metadata so the Navan entries match the actual dates, merchants, and totals.

The shown workflow includes a receipt upload from the desktop and an expense-details form with fields populated, including “Traveling: meals for myself” as the expense type and “Dominik Kundel” under participants. A final modal confirms: “Expense submitted. Your expense was submitted successfully.”

Kundel says Codex can upload missing receipts directly from the user’s computer. He also says the extension can leverage code execution, allowing Codex to control Chrome without the usual “screenshot, reason, move the mouse” loop. The claim is not only that Codex can see and click in a browser, but that it can script repetitive browser work where scripting is more efficient.

The extension’s parallel-tab behavior is shown with a multiplayer drawing game Kundel has been building. The prompt asks Codex to use four subagents to play the game with each other, with one Chrome tab for each agent and the first agent creating the game before the others join.

The browser shows multiple tabs assigned to different agents. A later view shows a four-panel grid in a room labeled “CODEX4,” with Agent 1, Agent 2, Agent 3, and Agent 4 drawing parts of a “tiny lighthouse” at the same time. Kundel says each agent has its own browser tab, so the agents can play in parallel and collaborate on the game.

That example presents the extension as more than a way to automate one foreground page. Codex can coordinate separate browser tabs for separate agents inside the same live web application. Kundel’s broader claim is that connecting Codex to the tools and apps the user already uses lets it do more work where that work is actually done.