TabCrew Online

TabCrew privacy policy

Effective date: November 2, 2026. Applies to the TabCrew browser extension and the tabcrew command-line tool (the local "daemon", and in hub mode the local "hub" started by tabcrew mcp), version 0.4.0 and later, and to the TabCrew Online relay at https://tabcrew.huynle.com (extension version 0.6.0 and later).

The short version

What TabCrew is

TabCrew lets AI agents that you run on your own computer (for example OpenCode or Claude Code) use your browser. It has two parts, both running only on your machine:

  1. The extension, which opens tabs for agents, reads those tabs, and clicks or types in them when an agent asks.

  2. The daemon, installed with npx tabcrew setup, which your agents connect to. It is free and open source.

    In hub mode (for browsers where native messaging is blocked), there is no installed daemon. Instead, your agent tool starts tabcrew mcp, which runs the same software as a "hub" while agents use it. In this mode the extension keeps the model provider settings and makes the model calls itself.

    In Online mode, the extension connects to the TabCrew relay instead, and your agents connect to the relay's MCP address (https://tabcrew.huynle.com/mcp). The software that runs on your computer in the other modes runs on the relay for you. See Online mode.

What the extension accesses, and why

Data When Where it goes
Text, links, buttons and form fields of pages in agent tabs When an agent calls observe, read_text, act or browse Only to the daemon on your computer
Screenshots of agent tabs When an agent calls screenshot, or calls browse while decision screenshots are on (see What the daemon sends off your computer) Only to the daemon on your computer; with decision screenshots on, the daemon forwards them to your Decisions provider
Console messages and request metadata of agent tabs (method, origin and path, status, type; never headers, cookies, bodies or query strings) When an agent calls debug_logs Only to the daemon on your computer
Titles, URLs, windows and tab group membership of tabs To keep each agent inside its own tabs and show sessions and groups in the popup Stays in the browser, or goes to the daemon on your computer
Which agent session owns which tab, and which tab groups and window are TabCrew's (tab, group and window ids, and session IDs) When agents open, finish or close tabs Stored in the extension's local storage on your computer, so agents stay isolated even after the extension reloads. Wiped when the browser restarts. Never sent anywhere else
The names of TabCrew tab groups (the default "TabCrew" group, groups you create or rename in the popup, groups agents ask for) When you or an agent create or rename a group Stored in the extension's local storage and shown as Chrome tab group titles; group names an agent uses are also kept with its session on the daemon or hub on your computer
The daemon's access token Fetched once from your computer by pairing (or typed by you) Stored in the extension's local storage on your computer
Your settings (tab placement, default group name, what happens to an agent's tabs when it finishes, connection mode, daemon or hub port) When you change them Stored in the extension's local storage on your computer
Hub mode: the pairing secret ("pin") shared with the hub Created on first connection Stored in the extension's local storage and in a private file on your computer (~/Library/Application Support/TabCrew/hub-pin on macOS). Never sent anywhere else
Hub mode: model provider settings (base URL, API key, optional headers, model names) When you click Save under Model provider Stored in the extension's local storage on your computer, not synced. The key and headers are sent only to that provider, with each model call. The popup never shows the saved key again (only its last 4 characters)
Hub mode: unsaved changes to the model provider form, including an API key you typed While you type, so they survive Chrome closing the popup for its permission prompt Kept in the extension's session storage (in memory, never synced) until you click Save or Remove, or the browser restarts (also cleared when the extension is reloaded). Never sent anywhere
Daemon mode: API keys and model names you enter under Advanced When you click Save Sent over the local connection to the daemon on your computer, which stores them in ~/.tabcrew/env or config.toml. The extension never stores them; the daemon shows only a key's last 4 characters back

Agents can only reach tabs they opened themselves, even when several agents share one tab group. Your own tabs and tab groups are never read or changed unless you give a tab to an agent, which TabCrew does not do on its own; a tab you drag into a TabCrew group stays yours. Password, file and hidden form fields are excluded from what agents observe.

Chrome may show TabCrew's tab groups as saved groups on the bookmarks bar. That is Chrome's own saved tab groups feature, stored by Chrome like any other tab group; TabCrew doesn't create bookmarks.

What the daemon sends off your computer

Daemon mode. Only one thing, and only if you use it: when an agent calls browse, the daemon sends the current agent page's visible text and its list of controls, plus the agent's goal and the steps taken so far, to the model providers you set up:

Screenshots (optional, off by default). If you turn them on (the popup's Attach a screenshot to each browse decision, decision_images in config.toml, or an agent calling browse with screenshot: true), the daemon also sends a screenshot of the visible part of the agent's tab (JPEG, at most 1568 pixels per side) with each request to your Decisions provider. Screenshots go to no other provider, and are never written to the browse result, to logs or anywhere on disk.

The daemon uses your own API keys, which are stored in ~/.tabcrew/env on your computer (whether you entered them during setup or in the extension popup), each sent only to the provider it is for: your OpenRouter key is never sent to a custom Decisions provider, which uses its own key (DECISIONS_API_KEY). Each provider's privacy policy applies to the data it receives. If you do not add a key, browse is disabled and nothing is sent.

tabcrew setup, tabcrew doctor --online and saving an OpenRouter key in the extension popup (or a Decisions key while the Decisions endpoint is OpenRouter's) also make one free request to OpenRouter to check that the key is valid. A key for another Decisions provider is not checked, so nothing is sent to that provider until browse runs.

What the extension sends off your computer in hub mode

In hub mode, the hub on your computer prepares the same browse requests. The extension sends them directly to the provider you saved under Model provider (for example OpenRouter, OpenAI or Anthropic), using your key and headers, and only to that provider's /chat/completions address. Chrome asks you to allow access to that provider's site the first time you click Load models, Test or Save, and nothing goes there until you allow it. Load models asks the same provider for its list of models (with the key and headers in the form). Test makes two small requests to it: it lists the models, and sends a tiny chat message. Once access is granted, opening the popup lists the models again by itself. If you set no provider, browse is disabled and nothing is sent. Hub mode does not use the Decisions API, so it never sends screenshots to a model provider. Nothing else changes: TabCrew still has no servers and no telemetry, and the developer receives nothing.

Online mode (TabCrew relay)

Online mode is for using your browsers from agents on other computers, or through an MCP gateway. You turn it on in the popup (Connection → Online), give the browser a name and sign in. The relay at https://tabcrew.huynle.com is run by the TabCrew developer on their own server.

Signing in. You sign in with GitHub or Google, in a browser window. TabCrew asks GitHub for your public profile and email address (read:user user:email), and Google for your basic profile and email address (openid email profile). It never sees your GitHub or Google password and never asks for access to anything else. The extension uses Chrome's identity permission only to show that sign-in window. Agents (MCP clients) sign in the same way, through the relay's sign-in page.

What the relay stores (in its database):

Data Why How long
Your account: the sign-in provider (GitHub or Google) and its user ID, your verified email address and display name, if the provider returns them To know which browsers and agents are yours Until your account is deleted
The names you give your browsers So agents can pick a browser, and so you can see and revoke them Until you revoke the browser or your account is deleted
The agents (MCP clients) you signed in: their name and the address they return to after sign-in So they can refresh their sign-in, and so you can see and revoke them Until revoked or deleted
Sign-in tokens, stored only as one-way hashes To keep the extension and your agents signed in Until they expire, are replaced or are revoked
One-time sign-in steps and codes To finish a sign-in 10 minutes and 1 minute
Your model settings, if you set them on the Model settings page: provider address, API key, optional extra headers and model names, and the optional Decisions provider (its address, API key and model, and whether to attach screenshots) So the relay can call your model and Decisions providers for browse from any of your browsers. Encrypted (AES-256-GCM) with a key only the relay server has; the API keys are never shown again (only their last 4 characters) and never sent to your browsers Until you click Remove or your account is deleted
Run recordings, only if you turn recording on: metrics, and with Full traces also page content per step (see Online mode, "Run recording") So you can see how your agents and browse perform, and review runs. Traces are encrypted (AES-256-GCM) with a key only the relay server has Metrics 90 days by default (30 or 365 to choose from), traces 14 days by default (7, 30 or 90 to choose from), or until you delete them or your account is deleted

What passes through the relay without being stored. Your agents' tool calls and their answers travel through the relay: the text, links, buttons and form fields of agent tabs, screenshots, console messages and request metadata, page URLs and titles, and the agent sessions themselves (labels, group names, which tabs belong to which session). The relay keeps them in memory only while they are in use. An agent session ends when the agent closes it, when you stop it in the popup, or after 30 minutes without use. The relay never writes screenshots (including those sent with browse decisions) or typed text to its database or its logs. It writes no page content to its database unless you turn on full traces (see Online mode, "Run recording"), and never to its logs. The connections are encrypted (HTTPS), but the relay itself handles this data unencrypted, as any MCP server does. If you add the relay to an MCP gateway, that gateway handles it too.

Model provider and browse. If you set a provider on the relay's Model settings page, the relay itself sends the browse requests (the agent's goal, and the current page's visible text and list of controls) to that provider, with your key and headers, and only to that provider's /models and /chat/completions addresses. Load models and Test on that page ask the same provider for its model list and send a tiny chat message.

If you also set a Decisions provider there (optional; the OpenRouter preset suggests https://openrouter.ai/api/alpha/decisions), the relay sends each step's decision request (the same goal, page text and controls) to that Decisions address instead, with the Decisions key, or with the provider's key only when the Decisions address is on the provider's own host. With Attach screenshots checked, or when an agent calls browse with screenshot: true, each Decisions request also carries a screenshot of the visible part of the agent's tab (JPEG, at most 1568 pixels per side), which the relay holds in memory only for that request. Screenshots go only to the Decisions provider. Save & test Decisions sends it one tiny question with no page content.

The relay only connects to providers, including Decisions providers, on the public internet over HTTPS. That provider's privacy policy applies to what it receives. If a browser uses its own provider instead (popup: Use this browser's own provider), the extension makes those calls directly, as in hub mode (no Decisions requests and no screenshots), and the relay never receives that key. Without either, browse is disabled and nothing is sent.

Run recording (optional). Recording is off by default. You turn it on for your account at https://tabcrew.huynle.com/account/recording. It has two levels:

You see your recordings at https://tabcrew.huynle.com/account/stats and https://tabcrew.huynle.com/account/runs. The developer, as operator, sees only totals across all accounts that turned recording on, never per-account rows, hostnames or traces. Metrics are kept 90 days by default (you can choose 30 or 365), traces 14 days by default (you can choose 7, 30 or 90). Turning traces off deletes them. The recording page has buttons to delete your traces or all your recordings. Deleting your account deletes them too. Daemon and hub mode have no recording.

Logs. The relay logs sign-in steps (for example "Google sign-in ok"), browser connections and disconnections, refused requests, and errors, with times. Logs never contain page content, typed text, tokens, codes or keys. The web server in front of the relay may log standard request details (IP address, time, address requested). Logs are kept on the developer's server and rotated.

Backups. The relay's database (the stored data above, including any recordings, still encrypted for traces) is backed up every night to storage the developer controls. Deleted data, including recordings, leaves the backups as they rotate.

Your controls. https://tabcrew.huynle.com/account lists your browsers and agents, with Revoke buttons; revoking one disconnects it and deletes its sign-in. https://tabcrew.huynle.com/account/recording turns recording on or off, sets how long it is kept, and deletes it. Sign out in the popup ends the browser's sign-in, and Pause agents in this browser stops agents from using it. To delete your account and everything stored for it, open an issue (see Contact) or use the support email shown on the Google sign-in screen; deletion is done within 30 days, and removed from backups as they rotate.

The developer does not look at the data passing through the relay or your recordings, except to investigate a problem you report, and never shares or sells it.

What is stored on your computer

Daemon mode: in ~/.tabcrew/ (readable only by your user account), the access token, your API key (if you added one), the list of open agent sessions (with their labels and group names), optional configuration, and service logs. Logs contain session IDs, tool names and timings; they never contain page text, screenshots, typed values, tokens or keys. npx tabcrew uninstall --purge deletes all of it.

Hub mode uses no ~/.tabcrew. It keeps the list of open agent sessions, its logs and its runtime files in a private per-user temporary folder, and the pairing pin in ~/Library/Application Support/TabCrew/ (macOS). The provider settings and key stay in the extension's local storage; unsaved changes to them stay in its session storage until you save or remove them, or the browser restarts.

In Online mode the extension keeps the relay address, the browser's name, its sign-in token (v2:online) and whether it uses its own provider in its local storage, and its own provider settings, if you set any, as in hub mode.

In every mode the extension's local storage also holds your settings, the TabCrew group names, and which tabs, groups and window belong to which agent session (ids only, wiped when the browser restarts).

Removing the extension deletes its local storage.

Sharing and selling

TabCrew does not sell, rent or share your data with anyone. In the local modes the developer does not collect it in the first place; in Online mode the relay keeps only the account data, model settings and, if you turn it on, the run recordings described above. Data is used only to do what you or your agents ask TabCrew to do in your browser.

Children

TabCrew is a developer tool and is not directed at children.

Changes

Changes to this policy are published in this file in the TabCrew repository, with a new effective date. The full history is public on GitHub.

Contact

Questions or concerns: open an issue at https://github.com/huynle/tabcrew/issues. For security problems, please use private vulnerability reporting as described in SECURITY.md.

TabCrew is an independent open-source project. It is not affiliated with, endorsed by, or sponsored by Google.