Give any AI agent a link to your app or your browser.

Most ways of letting an AI use your software tie you to one vendor or one machine. Agent Socket is a relay with one idea: your tab or app keeps an open line to it, and you give the agent a URL. Everything the agent does is an ordinary web request to that URL, answered by your tab or app as it happens.

• It's a link, not a plugin. Anything that can send an HTTPS request can use it: Claude Code on your laptop or in a cloud session, Codex, Cursor, OpenCode, an agent harness of your own, a twenty-line script. An agent running in the cloud can work in the browser on your laptop, signed in as you. Browser helpers built into one AI product work with that product, on that computer.
• Any web app can offer it. About ten lines of the SDK turn your app's actions into URLs an agent can call. There's no MCP server to host, no OAuth to set up and no backend of your own: the page itself answers. The Chrome extension is one app built this way.
• A link that holds. If your laptop sleeps, the Wi-Fi drops or Chrome restarts the extension, the relay keeps the session for up to 24 hours and the agent is told to retry. When your side comes back, the same link works again; you don't paste a new one.
• Skills agents share. A public registry of site profiles: notes and ready-made tools for a site, like reading Google Sheets' canvas grid. When you connect a tab, its site's profile loads, so the next agent doesn't spend ten minutes working the site out again. New profiles are reviewed before they're published.

Here is one call, from the moment your tab or app connects to the moment the agent has its answer.

1. Your side connects. The extension, or your app through the SDK, opens one WebSocket to the relay and lists the actions it offers. It gets back a link for this session.
2. You give the agent the link. Paste it into the chat. The agent fetches agents.md for instructions and tools.json for the list of actions; the relay answers both itself.
3. Every action is a web request. The agent calls an action like any URL. The relay passes it down the open line, your tab or app does it, and the answer comes straight back. A call can take up to 30 seconds; longer work replies with a task to poll.
4. You end it. Stop, closing the tab, or session.close() in your app ends the session and every link it made stops working.

The Chrome extension makes any tab you choose into an app on the relay. The agent can read the page, click, type, scroll, go to other pages and take screenshots, in that tab only, signed in as you.

1. Download the extension and unzip it.
2. Open chrome://extensions, turn on Developer mode, click Load unpacked and pick the folder.
3. On the extension's Details page, turn on Allow User Scripts. Without it, the agent can't run its own code on sites with a strict security policy.
4. Open the tab you want help with, click the Agent Socket icon, then Connect this tab.
5. Copy the link, paste it into your agent and say what you want done.

You stay in charge of what it can reach and for how long:

Give your app a “Connect with AI” button. Each tool is a path, a description and a handler; the handler runs in your page (or your Node process) when the agent calls it.

import { connect } from "https://esm.sh/@agent-socket/sdk@0.2"

let count = 0
const session = await connect({
  baseUrl: "https://agentsocket.dev",
  appId: "counter",
  agentsMd: "# Counter\nCall /increment to add to it.",
  tools: [{
    path: "/increment",
    description: "Add `by` (default 1). Returns the new count.",
    handler: ({ body }) => {
      count += JSON.parse(body || "{}").by ?? 1
      return { count }
    },
  }],
})
const link = await session.mintAgentToken({ label: "chat" })
// show link.url to the user: that is what they paste into the agent

The agent reads the link's agents.md, fetches tools.json, then sends POST …/increment with {"by": 2} and gets {"count": 2} back. The same app runs in Node. The limits are generous for this kind of work:

The full protocol, every limit and the threat model are in docs/protocol.md; the API in sdk/README.md.

Recorded on 2026-10-13 against the live relay: the demo app from “Try it on this page” running in Node, and a script in the agent's seat using nothing but web requests. Halfway through, the app's connection is cut without a goodbye, the way a sleeping laptop looks to the relay. Only the routine status checks are left out, and the link's secret part is hidden; nothing else is edited. The 200 replies are the demo app's own: it simulates Fig 1, so its words describe the switchboard.

The script that made this is relay/landing/capture.mjs.

A script with this page's link, in the agent's seat, drives Fig 1 through the live relay: answered calls, the laptop asleep and the line held, the plug pulled and calls turned away, then plugged back in. Recorded in a headless browser with no cuts, shown at twice the speed. The script is relay/landing/record.mjs.