knfcmbamhjmaonkfnjhldjedeobeafmk) · Last updated 2026-10-13The extension has no analytics or tracking backend. It communicates with the locally installed chrome-use CLI through Chrome native messaging. It handles browser observations and automation commands requested by the user. In the agent side panel, the selected local Codex, Claude Code, OpenCode or Pi CLI may send the user's task and requested browser context to its configured model service. That service's terms and privacy policy apply.
User-selected images, skill snapshots and scheduled prompts are saved locally with the profile or conversation. Image attachments and selected skill instructions are sent to the chosen agent/model service. Schedule and task-finish notification cursors store run, conversation and message IDs and timestamps, not prompts or replies. A finish notification shows at most the first line of the redacted answer. Imported skills are copied into a profile-owned library and are not executed by installation.
The side panel sends tasks to a local agent service through the native host. Conversation messages, replies and tool results are saved in a profile-scoped journal on the user's computer. The native service uses authenticated loopback IPC; the extension itself does not contact a model server. Windows uses protected user/SYSTEM ACLs for private state and owned process jobs for agent cleanup; macOS/Linux use private file modes and owned process groups. Chrome local storage keeps selection and message identifiers, not the conversation text. The panel does not ask you for an agent API key. Existing CLI authentication and provider configuration remain with the chosen CLI; discovery does not return credential values.
The current-page tag includes the active webpage's title, URL and tab identifier with the next message unless removed with ×. Page observations requested by the agent may also become model context. The model menu comes from CLI metadata; a custom model ID selects a model through that CLI.
Permissions depend on the agent: Codex/Claude use approval prompts or full access; OpenCode keeps CLI permission rules and can allow tools without asking; Pi exposes only chrome_use MCP tools in Browser only mode and its full tool set with full access. Full access automatically answers supported permission requests; it does not prevent task or page context from reaching the configured model service.
Deleting a stopped conversation removes its local panel history and browser session, plus Codex rollout files, Claude session files/directories, OpenCode history through session/delete, or Pi session files in the conversation folder. Codex's SQLite cache is not modified. Cleanup failures and unconfirmed browser-tab closure are reported.
| Data | Purpose and destination |
|---|---|
| Profile email | Read locally to label the connected Chrome profile and sent to the local native host. It is not used for advertising or analytics. |
| Task messages, replies and requested webpage content | Used to perform the user's browser task. The local agent may transmit task context to its configured model service. Messages and tool returns may be retained in the local conversation journal. |
| Target URLs and automation actions | Used to identify the intended page, execute requested actions and report their effects. Task-related URLs and tool returns may appear in local history or agent context. The extension does not harvest general browsing history or monitor unrelated user activity. |
| Cookies and authentication information | Cookie commands read or write the targeted site's cookies only when requested through the local CLI. The extension does not export cookies on its own or send them to a developer-operated server. The user's downstream CLI or agent controls any further handling. |
| Configuration and credentials | Discovery reads installation/configuration evidence. Panel edits are limited to supported model/provider fields. Credential values are not returned by configuration discovery or stored in the panel. The chosen CLI manages its own login. |
| Permission | Purpose |
|---|---|
| debugger | Attach the Chrome DevTools Protocol to tabs the automation session owns so the local CLI can read and drive them, only while the user is actively running a command. The user's other tabs are not attached. |
| tabs | Create, enumerate, and target the correct tab to automate, and route each command to it. |
| tabGroups | Group each session's own tabs into a labeled tab group so the user can visually distinguish agent-created tabs from their own. |
| downloads | Start a user-requested URL download, report its completion and local filename to the paired CLI, and list or clear Chrome download history. Clearing history never deletes files. Download metadata stays on-device and is sent only to the local CLI over native messaging. |
| nativeMessaging | The local transport to the paired chrome-use CLI — the extension's sole communication channel. |
| identity, identity.email | Read the email of the Google account signed into the current profile (chrome.identity.getProfileUserInfo) purely to label which profile/account a local connection belongs to, so a user with several logged-in Chrome profiles can confirm they are automating the intended account. The email is sent only to the local CLI over native messaging; it is never stored off-device or transmitted to any server. No OAuth token is requested (getAuthToken is not used). |
| sidePanel | Display packaged conversation, approval and local-agent configuration UI outside webpage DOM, available across tab navigation. |
| storage | Persist a per-profile random identifier and small local pairing/state values in chrome.storage.local. |
| alarms, idle | Keep the MV3 service worker alive during an active automation session and detect idle state so idle sessions can be cleaned up. |
| webNavigation | Detect when a driven tab finishes navigating so automation can wait for the right moment before reading or acting. |
| notifications | Surface local status notices (e.g. connection lost, a panel task finished) to the user. No data leaves the device. |
| activeTab, scripting | Inspect & Ask: only after the user presses its keyboard shortcut or the popup's "Inspect element" button, show an element picker in that tab, read the selected elements (text, sanitized HTML, styles, component names, source locations) and screenshot the visible tab for them. The result goes to the local native host only when the user presses Send, and from there through ocs to the agent session the user chose. Nothing is injected without that gesture and nothing stays in the page. |
We do not receive task contents through an extension-operated server and do not sell data or use it for advertising, credit decisions or unrelated purposes. Data accessed through the extension is used only for its browser-automation features requested by the user. A model service selected through the local CLI may receive the task context needed to perform that feature. Our use of information received from Google APIs adheres to the Chrome Web Store User Data Policy, including its Limited Use requirements.
Source code, issues, and contact: github.com/leeguooooo/chrome-use