Let your CellCog AI agent browse and act on the web using your real Chrome browser — your sessions, your bookmarks, your profile.
Browse my Chrome lets CellCog AI agents drive your real Chrome browser during a chat session. Instead of using a generic cloud browser, agents work inside your Chrome — using the sites you're already logged into, the cookies you've already set, and the bookmarks you've already saved.
Once enabled, an agent can:
Key Principle: You stay in control. The agent works inside a dedicated cyan "CellCog" tab group in your Chrome window — that's the shared workspace. Tabs outside that group stay private. Every action is classified by risk and gated by your approval threshold. Chrome itself shows a yellow safety banner on every controlled tab that you can dismiss to revoke access at any time.
Browse my Chrome is the peer of Cowork on my PC. Both require the CellCog Desktop app — they extend the agent's reach in two different directions:
| Use | Cowork on my PC | Browse my Chrome |
|---|---|---|
| Edit local files (code, docs, etc.) | ✅ | ❌ |
| Run terminal commands on your machine | ✅ | ❌ |
| Visit and interact with websites | ❌ | ✅ |
| Use logged-in sessions (Gmail, GitHub, banking) | ❌ | ✅ |
| Fill forms, click buttons, take page screenshots | ❌ | ✅ |
| Trust boundary | Working directory on disk | The "CellCog" tab group |
You can have both enabled in the same chat. Many real workflows use them together (e.g., "research these 8 SaaS products in Chrome, then write a comparison report into my project directory").
Browse my Chrome runs on top of the CellCog Desktop app — the same app used by Cowork on my PC.
/cowork/browse)If you already have CellCog Desktop installed and connected (from using Cowork), skip to Step 2.
For full Desktop setup details see the Cowork on my PC Guide.
The Chrome extension lives inside a specific Chrome profile. After installation, make sure that profile is open in Chrome so the extension can connect to the Desktop app.
You'll know it's working when:
Before enabling Browse in a chat, decide how much you trust the agent's self-classification.
On the Browse my Chrome page, find the Auto-approve threshold section and pick one of four levels — see Threat Levels and Auto-Approval below for details. The default is Safe, the strictest meaningful setting.
When you enable Browse, Cowork is also enabled automatically — Browse requires the Desktop app to be connected, and Cowork is how the chat talks to the Desktop app.
When the agent first opens a tab in Chrome, it creates a cyan tab group labeled "CellCog" in your Chrome tab strip. This group is the shared agent workspace:
If you already have a tab open that the agent should work with (e.g., your half-completed visa form), the agent can adopt it into the CellCog tab group. Because adopting necessarily moves your personal tab into the shared workspace, this action is classified as moderate and respects your threshold — at threshold Safe it requires your approval each time.
The agent perceives pages in three main ways:
If you have multiple Chrome profiles (e.g., Personal + Work), each one needs the extension installed separately. Once installed, each chat is bound to one specific profile via the Browse selector in the editor. The Browse my Chrome page shows the install state of every profile so you know exactly which one is missing the extension.
The agent uses Chrome's built-in chrome.debugger API (the same API used by Chrome DevTools and other AI-browser tools like OpenAI Codex and Browserbase) to navigate and act on pages. There's no separate browser launched and no special launch flag — it's your real Chrome doing real things.
When the agent reads a page, it gets:
When the agent takes an action, it can:
CellCog uses a transparent agent-classifies, user-threshold-gates model. The agent classifies every action by what data it touches; you control which classifications auto-approve.
Every browser action falls into exactly one of three risk levels, based on a simple question: what data does this command touch?
| Level | What it means | Examples |
|---|---|---|
| Safe | No private read, no data change. | Reading a Wikipedia article, navigating to a public URL, expanding a menu, searching on google.com |
| Moderate | Reads your private data but changes nothing. | Snapshotting Gmail / a private repo / banking, reading document.cookie, listing or adopting personal tabs |
| Dangerous | Changes data or triggers an irreversible action. | Submitting forms, sending email, paying, deleting, writing cookies, downloading files |
The agent shows you both the classification AND a one-sentence reason on the approval card. You're always picking how much to trust the agent's own self-classification — not picking a category of content the system identifies.
The Browse my Chrome page has four threshold levels. Anything above your threshold still requires your manual approval on the approval card in chat.
| Threshold | What it auto-approves |
|---|---|
| None | Nothing — every browser action requires your approval. The strictest setting. |
| Safe (default) | Side-effect-free actions on public content. Reading your private pages and any data changes still require approval. |
| Moderate | Also auto-approves actions that read your private data — logged-in page snapshots, cookies, personal-tab discovery. Anything that changes data still requires approval. |
| Dangerous | Auto-approves everything the agent declares — including actions that change data. The most permissive setting. |
Most users settle into Safe or Moderate. Dangerous is intended for advanced workflows where the user trusts a specific agent role completely (e.g., a project-scoped agent reviewing their own work) — not a default.
Chrome shows a yellow banner across the top of every tab CellCog is driving:
"CellCog Browse started debugging this browser." — Cancel
This is Chrome's built-in safety indicator. CellCog cannot hide or suppress it. Click "Cancel" at any time to immediately revoke debugger access for that tab — the agent will see an error on its next attempt and can prompt you to re-enable in the Browse my Chrome settings.
You have multiple escape hatches, at different scopes:
chrome://extensions/When Browse is enabled for a chat:
Browser_Tabs list, which is classified moderate)The extension automatically redacts before any data leaves Chrome:
type=password)autocomplete=cc-number)autocomplete=cc-csc)These values are replaced with <redacted> in everything the agent reads.
These are not auto-redacted because they sometimes legitimately need to flow through the agent (e.g., the agent reads a session cookie to make an authenticated API call on your behalf):
For full disclosure of these data flows, see the Privacy Policy.
/cowork for a green "Connected" status.chrome://extensions/, find CellCog Browse, click the reload icon. Then revisit the Browse my Chrome page.This usually means the chat is bound to a different Chrome profile than the one with the extension installed.
The yellow banner is Chrome's built-in indicator and only goes away in two situations:
If neither has happened and the banner persists, check that the CellCog Desktop app is still running. If Desktop quits mid-session, the debugger may stay attached until you reload that tab.
The agent tried to act on a tab that hasn't been adopted into its workspace. Either:
Browser and goto)When you click "Install →" in CellCog, Chrome opens new tabs in whatever profile is currently focused — not necessarily the profile you intended. If the extension landed in the wrong profile:
chrome://extensions/The Desktop app is the trust boundary — it authenticates you and brokers every command to and from CellCog's servers. The Chrome extension is the small piece that talks to Chrome on the Desktop app's behalf. Splitting them keeps the powerful pieces (the Desktop app) under your direct control, and keeps the extension intentionally minimal.
Not currently. Browse my Chrome supports Google Chrome only at launch. Microsoft Edge and other Chromium-based browsers are on the roadmap; Firefox and Safari are not currently planned because they don't expose an equivalent to chrome.debugger.
The agent can list them (only when you grant approval — Browser_Tabs list is classified moderate and requires your threshold to be ≥ moderate to auto-approve). It cannot interact with them. To interact with one, the agent must adopt it, which moves the tab into the CellCog group and is a separate, also-moderate action.
No. Password fields and credit-card fields (number + CSC) are always redacted before any data leaves Chrome. The agent sees <redacted> in those fields and cannot retrieve the real values.
Screenshots are saved to a temporary directory on your machine and made available to the chat session (so the AI can analyze them). They're transmitted to your chosen LLM provider as part of the chat context. They're not stored permanently by CellCog or shared elsewhere.
You can use Browse my Chrome on any site you like. Snapshotting a banking dashboard is classified moderate (reads private data, changes nothing); clicking "Transfer $1000" would be dangerous. Your threshold setting decides what runs without your click — most users keep the threshold at Safe or Moderate when working with sensitive surfaces.
No. Browse my Chrome is included with CellCog. Standard credit usage applies to the agent operations the same way Cowork does — credits are consumed by the agent's reasoning and tool calls, not by the act of having Browse enabled.
No. Browse runs against your Chrome profile on your machine, authenticated as you. Organization members can't access your machine through CellCog. If they want to use Browse, they install the Desktop app and the extension on their own machine.