André Bandarra

Developer Relations Engineer for Chrome and the Web at @Google, technology addict.

[Sarah Drasner @sarah_edo]

Really enjoying this talk and panel from @domfarolino @liadyosef and @ochafik at AGNTCon!

They’re not just talking about WebMCP or MCP Apps, but also how they can interact together!

If you’re not familiar with MCP Apps, here’s some documentation

https://t.co/EMCrDCzb0v

[Oscar Falmer @OscarFalmer]

I am very excited about the new WebMCP feature we announced at Meta Connect!

It allows Meta AI to seamlessly interact with your web app on your glasses, perfectly bridging the user, the AI agent, and your web app.

Want to learn more? Here is the presentation by @_minglin_

[André Bandarra @andreban]

We're going live in a now :D. Join us!

[Kasper @kaskulikowski]

🎬 Next on The Agentic Web livestream: WebMCP Skills

We have new experimental skills in the webmcp-tools repo 🧪

Join @andreban and me live to see them design, test and build WebMCP tools!

📺 https://t.co/qMfXctTZIv , Friday Oct 23rd, 5pm CET, 8am PDT

[André Bandarra @andreban]

@SeraAndroid @NVIDIAAI 🤩

[André Bandarra @andreban]

@cheneytsai @ndmccormack @OpenAI 👀

[André Bandarra @andreban]

@b1tr0t It doesn't solve the gotcha that you can't throw any error and both end up being quite similar, as you'd throw a ToolCallError in exactly the same places you'd return and object with isError, but now you have different treatment for different error types.

[André Bandarra @andreban]

⚠️ A #WebMCP gotcha worth knowing about: throwing errors inside native registerTool() handlers.

The browser hides your message and the agent only sees "DOMException: UnknownError". Return the error as a result instead, and the agent can fix its own call.

[André Bandarra @andreban]

@b1tr0t Some of this comes to tool design, and user generated content can be isolated in a separate tool. This design makes a lot of sense for things like product reviews, for instance. The freeform notes is tricky indeed, and a per-fiels hint could help.

[André Bandarra @andreban]

🔒 "It's our own database, so it's safe" is a very natural thing to assume. With #WebMCP agents reading that data, it's worth a second look.

WebMCP has a hint for this case, and JupyterLite is a nice example of using it.

[André Bandarra @andreban]

@0xidanlevin LOL

[Niall McCormack @ndmccormack]

I'm excited to share that WebMCP is now enabled for MCP Apps in ChatGPT Desktop! 🚀

@OpenAI is also collaborating with https://t.co/WAjlykp7dX to make WebMCP the default way for agents to interact with MCP Apps.

Another big step toward a more seamless agentic web!

[André Bandarra @andreban]

@kaskulikowski So, an MCP App returned via a WebMCP tool on the page, and then the agent talks to the MCP App via WebMCP?

[André Bandarra @andreban]

JupyterLite WebMCP: https://t.co/Skf7zmNg50

[André Bandarra @andreban]

The hint doesn't make prompt injection impossible on its own. But It gives the agent the signal it needs to treat the output as data.

[André Bandarra @andreban]

Set untrustedContentHint: true on any tool that returns user-written text.

That signals the host agent that the output came from an untrusted source, so the host can isolate it as data instead of treating it as instructions.

JupyterLite sets the hint on all 19 tools that read notebook cells.

[André Bandarra @andreban]

@madmanstudioio Yeah, when things go wrong you can craft a message and help the agent recover. UnknownError doesn't do that.

[André Bandarra @andreban]

@b1tr0t Would this mean something like catch an error, then re-throwing as something like a ToolCallError? That is, you'd still need the catch / return something else flow, but in this case a different error?

[André Bandarra @andreban]

Shout out to @sebasmediaprod, creator of Roque Nights!

[André Bandarra @andreban]

JupyterLite WebMCP: https://t.co/Skf7zmNg50

Roque Nights: https://t.co/YFFOnprh8m

[André Bandarra @andreban]

@sebasmediaprod @OpenAI Congratulations, and thanks for creating such a great app!

I've been analyzing what the Hackathon winners did well and just mentioned Roque Nights as a good example of handling errors here https://t.co/tr6WEiaJba

[André Bandarra @andreban]

Two hackathon winners that do this well:

📓 JupyterLite WebMCP returns STALE_CELL with the cell's new content, so the agent can retry without another read.

🔭 Roque Nights puts suggestions in its hints: "Target not found. Did you mean Messier 31?"

[André Bandarra @andreban]

@nartmadi @kaskulikowski This one doesn't depend. :D