TabPilot app icon

TabPilot

MCP for Safari.

Let Claude Code, Codex or any MCP agent use your real Safari — your tabs, your logins — and ask you before anything risky.

Coming soon to the Mac App StoreSee how it works ›

Runs locally on 127.0.0.1 · No account · Requires macOS 14 or later

Works with Claude Code, Codex and other MCP clients over HTTP. Claude Desktop connects through a one-line bridge.

Three steps. One command.

Install it, switch it on in Safari, and tell your agent where to find it.

  1. Install TabPilot

    Get it from the Mac App Store and open it once. It starts a small server on this Mac.

  2. Turn on the extension

    Safari › Settings › Extensions › TabPilot. Allow it on the websites you want your agent to use.

  3. Paste one command

    TabPilot shows it with your private token filled in. Your agent now has the tools.

Show the setup for
The app shows it with your token filled in
claude mcp add --transport http tabpilot http://127.0.0.1:47823/mcp \
  --header "Authorization: Bearer <token>"

Run it once in Terminal. The token works only on this Mac; anyone who has it can control your Safari, so keep it private.

{
  "mcpServers": {
    "tabpilot": {
      "command": "npx",
      "args": ["-y", "mcp-remote", "http://127.0.0.1:47823/mcp", "--header", "Authorization:Bearer <token>"]
    }
  }
}

Needs Node.js. Claude Desktop only starts local programs, so this goes through the mcp-remote bridge. Add it to claude_desktop_config.json and restart Claude.

URL:     http://127.0.0.1:47823/mcp   (Streamable HTTP)
Header:  Authorization: Bearer <token>

Any MCP client that speaks Streamable HTTP. The server only listens on this Mac (127.0.0.1).

Step by step: connect Claude Code to Safari · Claude Desktop and other clients

Your Safari. Your rules.

The agent gets real tabs and real logins. You keep the final say.

Site Access

Ask First, Read Only, Read & Act or Off, set per website. New sites ask before the agent sees anything.

Approve risky actions

Buy, pay, send, delete and submit wait for you to choose Allow Once or Don’t Allow.

Secrets stay yours

Password, card and one-time-code fields are left out when the agent reads a page, and it can’t type into them. They carry a “Hidden from agent” badge.

Prompt-injection shield

Hidden text is removed and instruction-like lines are marked as untrusted page content.

Take Over

For a sign-in, 2FA or a captcha, the agent hands the tab to you. Press Done and it carries on.

Agent cursor

A blue cursor, a focus ring and a caption show exactly what the agent is about to do.

Stop button

“TabPilot is controlling this tab · Stop” sits in the page. Or stop from the toolbar, or press Option-Shift-P.

Send to Agent

Right-click a selection, a link, an image or a page. It waits in Notes for your agent’s next step.

Activity log

Every read, click and prompt, with the site and the time, in the toolbar popover and the app.

Local only

The server listens on 127.0.0.1 behind a token. No account, and no cloud of ours.

It asks before anything risky.

When the agent reaches for Pay, Send or Delete, TabPilot stops and asks — right in the page, the way Safari asks for permissions.

Site Access, per website

Ask FirstThe agent asks before it can see this site. The default for new sites.
Read OnlyReads pages, never clicks or types.
Read & ActReads, clicks and types. Risky actions still ask.
OffThe agent can’t see the site.

Local only. By design.

TabPilot runs a small server on 127.0.0.1, protected by a token only your agent knows. There’s no account and no cloud of ours. Pages you let the agent read go to the agent you connected — nowhere else.

127.0.0.1 onlyNot reachable from other computers.
Token requiredMake a new one anytime; old setups stop working.
No accountNothing to sign up for. No analytics.

Privacy policy

Not a clean automation window.

FeatureTabPilotAutomation window
Your logins and open tabsYesNo, starts from a clean slate
Asks before risky actionsYesNo
Per-site access rulesYesNo
Best forYour everyday errandsTesting your own website

Safari 27 has its own MCP server for web developers, which works in automation windows. Safari’s MCP server vs TabPilot

Questions

What is MCP?

The Model Context Protocol is how AI agents get tools. TabPilot gives your agent Safari tools: list tabs, read a page, click, type, take a screenshot and ask you for help.

Which agents work?

Claude Code, Codex and other clients that support MCP over HTTP. Claude Desktop connects through a one-line bridge that TabPilot shows you. Setup for Claude Desktop and other clients

Can the agent buy something without me?

Buy, pay, send, delete and submit always wait for Allow Once. Sites set to Read Only can’t be clicked at all.

Does it see my passwords?

No. Password, card and one-time-code fields are left out when the agent reads a page, and it can’t type into them. For a sign-in, it asks you to Take Over.

What about prompt injection?

Hidden text is removed and instruction-like lines are marked as page content before the agent reads them. That lowers the risk; it can’t remove it. Keep sensitive sites on Read Only or Off. How to let an agent use your browser safely

How is this different from Safari’s own MCP server?

Safari 27’s MCP server is made for web developers and works in clean automation windows without your logins. TabPilot works in your everyday Safari. The comparison

Does the app need to stay open?

Yes. The TabPilot app runs the local server. Quit it, and the agent’s tools stop working.

What do I need?

A Mac with macOS 14 Sonoma or later, Safari, and an MCP client such as Claude Code. TabPilot is a one-time purchase on the Mac App Store, with no account and no subscription.

Is there a safe page to try it on?

Yes: the TabPilot Demo Store. Nothing on it is real, and nothing you type leaves the page.

More help: Support · Guides