In-browser AI agents (WebMCP)

Yes. When you pair a tab, tick Also allow in-browser AI agents (WebMCP), and an AI agent built into your browser can use that tab's DiceyTable tools. It follows the same rules as the agent in your AI app.

WebMCP is a new web standard, still in early trials. It lets a web page offer tools to an AI agent that lives in the browser itself or in a browser extension. Regular MCP works differently: your AI app starts a helper program, the DiceyTable bridge, and talks to it. What is MCP? explains the regular kind.

What WebMCP tools does every page offer?#

Every DiceyTable page offers five small tools to any WebMCP agent, with no pairing and nothing to set up. They are anonymous. They read the same public information any visitor sees, and they never use your account, a table or the editor.

Game names and descriptions are written by game authors, so these results carry the untrusted-content flag too. To build, edit or play games, the agent needs the paired tools below.

How to turn it on#

WebMCP rides along with a normal pairing, so you still need the bridge first.

1. Ask your agent to connect, and open Connect AI agent in your tab. The steps are in Pair a DiceyTable tab with your agent.
2. In the consent dialog, tick Also allow in-browser AI agents (WebMCP). It starts off.
3. Type the code and click Allow.

Once the tab is paired, DiceyTable starts a second agent session just for the browser's agent, signed in as you and named In-browser AI agent. Then it offers the tab's tools to the browser. The list changes as you move around: the editor offers editor tools, a table offers table tools. If your browser has no WebMCP of its own, DiceyTable loads a small stand-in so that extensions which speak WebMCP can still find the tools.

You don't need to switch on any browser flag for the bridge. Whether your browser's own agent needs one is up to that browser.

The same rules apply#

Every call from the browser's agent goes through the same checks as a call from the bridge:

• the permissions on your key, and never more;
• the same rate limits, and the same Destructive edits choice you made;
• the same actions that need your click, such as publishing;
• the same Activity… list, so you can see every call.

Other players see the AI badge, and the usual rules about dice chips and stats apply too. Tools that return text from outside you, like card text or chat, carry WebMCP's untrusted-content flag. That tells the browser's agent to treat the text as data, not orders. See Prompt injection protection for AI agents.

What can't an in-browser agent do?#

• Import files. File bytes reach the tab only through the bridge, so asset_import doesn't work here.
• Ask to publish. Its session never gets the Request publishing permission, even when your key has it.
• Use the session_* tools. Those live in the bridge, not the page.

Screenshots work, but WebMCP has no picture type. An image comes back as text-encoded data inside the answer, which is clumsy for most agents. Use the bridge when you want screenshots.

How do I turn it off?#

The browser's agent stops when the tab's pairing ends. Click Disconnect on the Agent connected chip, or end the session in Settings ▸ AI agents. To keep the bridge but drop the browser's agent, disconnect, then pair again with the box unticked. See Watch, pause or stop your agent.

Next steps#

• How the DiceyTable MCP bridge works: the bridge, the tab and the relay.
• Is it safe to let an AI agent use my account?: the full safety picture.