Skip to main content

Universal AI Agent Connector — Privacy Policy

Last reviewed: 2026-11-09. Applies to Universal AI Agent Connector 0.1.0 (Chrome extension) and its local daemon (@corsen-ai/universal-connector, distributed as source at github.com/CorsenAI/universal-ai-agent-chrome-connector).

Universal AI Agent Connector links the Chrome tabs you choose to the AI agents that run on your own computer. This policy states what the extension handles, where it goes, and what it does not do.

Publisher: Corsen AI. Contact: hello@corsen.ai.

The short version

The connector itself runs only on your computer. Corsen AI operates no server, relay, account system, analytics service or telemetry endpoint for it, and receives none of the data described below.

The extension talks only to the connector's daemon on 127.0.0.1, on your computer. The agents you connect to that daemon receive what they ask for from the tabs you chose, and send it wherever you configured them to send it — typically their model provider. That part is governed by those agents' settings and policies.

Data the extension handles

Content of tabs you expose to agents (role "MCP")

What. For a tab you gave the MCP role: its visible text (as Markdown or as an accessibility tree with short element handles), title and address; console errors and warnings, JavaScript errors and failed network requests; the JSON and GraphQL request and response bodies the page itself exchanges; and a screenshot when an agent asks for one.

When. Only for a tab you gave this role, and only while it keeps it. Page reads, actions and screenshots happen only when one of your agents calls a tool; console and network capture runs continuously for an exposed tab. Every tab starts with no role.

Where it goes. From the tab, to the extension, to the daemon on 127.0.0.1, to the agent that asked. The extension adds no other destination.

Masked before it leaves the page:

  • password fields and hidden form fields;
  • fields the page declares as one-time codes, passwords or payment card data;
  • fields whose name or label indicates a password, secret, token, API key, one-time code, card number, security code, IBAN, SSN or PIN;
  • credential-bearing parameters in addresses, and any user name or password embedded in an address;
  • in captured network traffic: credential headers (authorization, cookie, set-cookie, x-api-key and similar) and secret-looking keys in request and response bodies.

Honest limitation. Detection relies on the signals a page gives; a field with no telling type, name or label cannot be told apart from an ordinary one. A screenshot is pixels and nothing in it is masked: it shows the page as you see it.

Chat tabs you use as a model (role "LLM")

What. For a tab you gave the LLM role, the prompt a tool on your computer sends to the daemon is typed into the chat page, and the newest answer is read back from the page as text.

Where it goes. Between the daemon, the extension and that chat page. The chat service receives the prompt as if you had typed it, under your account and its terms. If you calibrate a chat site, the extension stores the CSS selectors you picked for it — never page content.

Tab titles and addresses

The side panel lists your open tabs locally so you can choose roles. Only tabs that have a role are reported to the daemon (title, address with credentials removed, error counters); for every other tab, only their number.

Settings

The pairing code (used to compute proofs for the daemon, never transmitted), a random identifier for this installation, the daemon's port and your calibrated selectors are kept in this browser profile's extension storage. Tab roles are kept for the current browser session only.

What the connector does not do

  • Does not send anything to Corsen AI. No analytics, telemetry, crash reports or usage statistics.
  • Does not read your browser's cookie jar, history, bookmarks, saved passwords or other extensions' data. Cookie headers in captured traffic are masked before leaving the page.
  • Does not read or act in a tab without a role. No debugger is attached to it, nothing is read from it, no input reaches it.
  • Does not act on its own. It acts when you ask it to from the side panel or dashboard, or when a tool on your computer that holds the daemon's key asks it to, and only in tabs you chose. Agents cannot change a tab's role.
  • Does not write page content to disk. Content is produced on request and forwarded, or held in bounded in-memory buffers for exposed tabs (console entries, network calls and bodies); removing the tab's role or closing the tab clears them. The daemon keeps an activity summary (tool names, timings, sizes), not content.
  • Does not sell or transfer data, and does not use it for advertising, profiling or creditworthiness.

The local daemon

The daemon listens on 127.0.0.1 only and refuses requests from web pages (Host and Origin checks) and requests without its key. It stores its pairing code and API key in ~/.corsen-connector/credentials.json (%USERPROFILE%\.corsen-connector\credentials.json on Windows), readable by your user account only (file mode 0600 on POSIX, the profile's access control on Windows). When you connect an agent, it writes a timestamped backup of that agent's configuration next to the file (.corsen-backup-….bak) before changing it.

Your control

  • Panic stop (side panel or dashboard) detaches every tab and resets all roles. Cancelling Chrome's "started debugging this browser" bar makes a tab private again.
  • universal-connector token --rotate replaces the pairing code and the API key; universal-connector unlink --all removes the connector from your agents.
  • Removing the extension deletes its storage. Deleting ~/.corsen-connector (%USERPROFILE%\.corsen-connector on Windows) removes the daemon's own state; the .bak copies it wrote next to your agents' configuration files remain — delete them too for a clean removal.

Limited Use

The use of information received from Chrome APIs adheres to the Chrome Web Store User Data Policy, including the Limited Use requirements.

Changes

Changes to this policy are dated here. A change in how data is handled is announced in the release notes before it ships.

Contact

Corsen AI — hello@corsen.ai. Security reports: see SECURITY.md in the repository.