AgentHub — Privacy Policy

AgentHub is a Chrome extension that connects your browser to AI assistants (an in-extension chat agent, or external AI tools via the Model Context Protocol). This policy explains what data AgentHub handles and what it does not.

1. Data AgentHub processes

When you invoke an AI action (via the chat sidebar or an MCP-connected tool), AgentHub reads data from the active tab to fulfill that action:

• Page content (text, HTML, structured data)
• Page metadata (title, URL of the active tab)
• Screenshots (when you or the AI explicitly takes one)
• Form field contents (when the AI is asked to read or fill a form)
• Open-tab list (when the AI is asked to list or switch tabs)

This data is only collected at the moment you invoke an action, and is sent to the AI provider you've configured. It is never sent to us. Separately, and independently of any AI action, the local bridge keeps diagnostic logs and — in packaged builds, unless you opt out — uploads a redacted form of them to us; see section 2.

2. Where your data goes

Chat sidebar (Phase 1, user-supplied API key)

• Chat messages and the tool-result snippets the AI needs to answer are sent to api.openai.com over HTTPS, using your own OpenAI API key, which you paste into Settings and which is stored only in your browser's local extension storage. AgentHub's developer never sees this key or the data it transmits.
• See OpenAI's privacy policy for how OpenAI handles requests.

MCP server (external AI clients)

• When you connect an MCP-capable AI tool (Claude Code, Cursor, VS Code, Windsurf, Continue, JetBrains, Zed, etc.), tool calls flow locally between that tool and AgentHub's local native-messaging host over localhost — no internet hop.
• Page data returned to your AI tool is then transmitted by that AI tool to its own provider (e.g., Anthropic for Claude Code, your configured model provider for Cursor) according to that tool's privacy policy.

Diagnostic logs — the one thing that is on by default

The local bridge writes diagnostic logs (tool-call metadata, errors, timings). In packaged builds — the binary the setup command installs — a redacted copy of those logs is uploaded to an ingest endpoint we operate, batched and fire-and-forget. This is opt-out, not opt-in: it is on unless you turn it off.

• What is redacted before upload: page content, form values, cookies and secrets are replaced with structural summaries (for example [len=42] or [REDACTED-SECRET]) rather than their values. URLs are reduced to scheme and host — the path and query are dropped.
• How it is keyed: a random per-install UUID, generated once and stored next to the bridge. It identifies an install so logs correlate across restarts — it is not derived from you, your hardware, or your browser identity.
• How to turn it off: set remote.enabled to false in logs-config.json in the AgentHub install directory. Nothing is uploaded after that. Setting the top-level enabled to false disables all logging, local files included.
• Where it does not apply: the dev/test bridge run from source has no baked-in endpoint and uploads nothing unless a config explicitly opts in.

3. Data stored on your device

The following are stored in Chrome's local extension storage (chrome.storage.local) — never transmitted off-device by AgentHub. (The bridge's own diagnostic logs are separate and are covered in section 2.)

• Your OpenAI API key (for chat)
• Per-tool permission toggles
• Domain blocklist (sites where the AI is not allowed to act)
• Connection state for the MCP transport
• Real-time activity log of AI actions

You can clear all of this at any time by removing the extension or using your browser's site/storage controls.

4. Permissions and why we ask for them

• tabs — list/switch/target tabs when your AI requests it
• sidePanel — render the chat, tools, and settings UI
• nativeMessaging — talk to the local MCP server binary
• scripting — run content scripts on the active tab during AI-invoked actions
• storage — persist preferences and API key locally
• debugger — Playwright-based form filling (iframes, ARIA widgets, contenteditable, file upload, date pickers)
• alarms — periodic heartbeat for the MCP connection
• <all_urls> (optional) — granted only if you opt in via Settings; lets the AI act on any site you visit. Default is per-site (you grant each site individually).

5. What AgentHub never does

• Never sells your data to third parties.
• Never uses your data to train AI models.
• Never transmits page content to AgentHub-operated servers. (We do operate one ingest endpoint, for the redacted diagnostic logs described in section 2 — page content is stripped before upload.)
• Never reads passwords, payment information, or authentication tokens.
• Never operates on a site that's in the default blocklist (banking, email) unless you explicitly remove that site from the blocklist.

6. Children's privacy

AgentHub is not directed at children under 13. We do not knowingly collect data from children.

7. Changes to this policy

Material changes will be reflected by updating the "Last updated" date above and (where appropriate) notified inside the extension before they take effect.