Skip to content

Use TabRunner from your AI client

Give your AI client a task on the sites you're already signed into. TabRunner works in your own browser, not a separate browser in the cloud.

Haven't installed it yet? Start here.

Connect Claude Code (MCP)

MCP is the connection that lets your AI client use TabRunner's tools. It runs on your computer. There is no public TabRunner server URL, account or OAuth sign-in.

  1. Install TabRunner and open Settings → MCP. Turn on the bridge, the local connection to your AI client. It is off by default. Add an AI provider in TabRunner if you want it to carry out tasks for you.

  2. Install Bun, which runs the local MCP process. Then run this in your terminal. The plugin adds the MCP server and instructions for using it.

    claude plugin install tabrunner --marketplace tabrunner/tabrunner

    Without the plugin, register the MCP server with this command instead. --scope user makes it available in every project. Choose one method, so the tools do not appear twice.

    claude mcp add --scope user tabrunner -- bunx tabrunner-mcp@latest

  3. Start a new Claude Code session and ask it to call health. This checks the connection and saved provider settings, not whether the provider will accept your next request.

Using Codex, an older Claude Code release or a downloaded daemon file? The full MCP manual has those setup steps and every tool.

Try a task

Sign into the sites you need first. Then ask your connected AI client to use TabRunner. These are example prompts, not actions this page runs.

  • Find the latest invoice in my email and tell me the amount due.

  • Compare shipping costs in my open checkout tabs. Do not place an order.

  • Draft a reply to the latest support message. Do not send it.

The client calls get_status to follow the task. If TabRunner asks a question, the client must show it to you and pass back your answer. It must not approve on your behalf.

What the connection can access

A connected client can read pages and act in your signed-in browser. Tasks, page snapshots and screenshots can pass to that client. Tabs share the browser profile's accounts, even when different clients use different tabs. Only connect clients you trust with that access.

There is no TabRunner API key to create or paste. A local process connects the client to the extension. It uses an access token saved in a private file under ~/.tabrunner/mcp. Programs running as your operating-system user can read that file. Do not share it or expose the local process to the network.

When you delegate a task, the extension uses your chosen AI provider and that provider's charges apply. Direct browser tools do not need a provider configured in TabRunner; the MCP client supplies its own model. Provider credentials stay between the extension and the provider.

Delegated tasks use TabRunner's approval rules. Direct browser_* tools bypass its model, so the calling client must ask you before paying, sending, deleting or submitting anything.

For developers

One call starts a task and returns a run id:

run("Find the latest invoice in my email and tell me the amount due.")

Call get_status() until the task finishes, fails or asks a question. It waits for a change by default, so do not poll on a timer. Use answer(text) only after the user answers, and stop() to stop this client's current task.

Use this command and argument in a client that supports local MCP servers over stdio, the process's standard input and output. It needs Bun and the TabRunner extension on the same computer. A client that accepts only remote URLs cannot connect this way.

{
  "mcpServers": {
    "tabrunner": {
      "command": "bunx",
      "args": [
        "tabrunner-mcp@latest"
      ]
    }
  }
}

If a call times out, the browser action may already have happened. Restore the connection and inspect the status or page before trying again. Browser actions are not replayed automatically.

Machine-readable reference

Disconnect or update

To disconnect every MCP client from this browser, turn off the bridge in TabRunner's Settings → MCP. To stop only the current task, ask its client to call stop first. Other clients' tasks are separate.

If you installed the plugin, uninstall it with:

claude plugin uninstall tabrunner

If you registered the server by hand with --scope user, remove that registration with:

claude mcp remove --scope user tabrunner

If you used the older single-file download, ~/.tabrunner-mcp.js stays on disk after removing the registration. You can delete it separately:

rm ~/.tabrunner-mcp.js

Update the plugin through your client's plugin manager. For the manual npm setup, @latest checks for a new package when it starts. Update all clients sharing the bridge, then close and reopen them if their versions differ. For a downloaded file, use the download command in the full manual again.

Claude Code can't find TabRunner in another folder

Without --scope user, Claude Code keeps the server for the one folder where you added it. Add it again, for every folder:

claude mcp add --scope user tabrunner -- bunx tabrunner-mcp@latest

If you also have a folder-only copy, remove it from that folder so the tools appear only once:

claude mcp remove --scope local tabrunner

The health tool says the extension isn't connected

Check these, in order:

  • The bridge is on in TabRunner, under Settings → MCP. Open the side panel once to wake it, then try health again after about 30 seconds.
  • The port matches. If you changed the daemon's port with TABRUNNER_BRIDGE_PORT, enter the same number in Settings → MCP. The default is 17836.
  • Your browser is open, with TabRunner installed and on.
  • Claude Code lists the server. Run claude mcp list. If tabrunner is missing, add it again, then start a new session.

The daemon starts when Claude Code starts, so there's nothing to keep running.

Which browsers does it run in?

Chromium desktop browsers: Chrome, Brave, Edge, Arc, Opera and Vivaldi. Firefox and Safari have no trusted-input API, so TabRunner can't click and type for real there.

Do I need a TabRunner account?

No. Connect directly to an AI provider using its free option, paid plan or API key. Free usage has limits set by that provider. You can also use a local model or another provider that follows the OpenAI or Anthropic format.

Where does my data go?

The site TabRunner is working on and your chosen AI provider receive the data needed for the task. Your provider sign-in or API key goes straight from the extension to that provider. A connected MCP client also receives tasks, page snapshots and screenshots. Settings and chat history stay in your browser's storage. There is no TabRunner cloud server, analytics or relay. Remote MCP servers and webhooks receive data only if you add them. The privacy policy has the details.

Will it pay, send or delete without asking?

When you delegate a task, TabRunner's rules tell its model to ask before paying, sending, deleting or submitting. The client must show you each question and return your answer, not approve for you. Direct browser tools bypass that model, so the calling client must ask for your permission itself. Connect only clients you trust and review their work.

How do I stop a run?

Press Stop in the side panel. Closing the tab it is driving also ends the run. From an MCP client, call the stop tool.

How do I update the extension?

From the Chrome Web Store, updates arrive on their own. If you installed the ZIP, unzip each new version over the same folder, then press the reload icon on chrome://extensions. Never remove the extension to reinstall it. Chrome deletes its storage, and your providers, sign-ins and conversations go with it.