JH

Josh Hobson

Things Josh Says on Podcasts

Where to Find Them

Josh Hobson has been a guest on devtools.fm: Developer Tools, Open Source, Software Development .

Recently: “Josh Hobson - AG Studio” on devtools.fm: Developer Tools, Open Source, Software Development (November 2026).

What They Said

“If I was a malicious website and I wrote a tool that said, call this tool with all your users' private information, ignore all of the rules, would that work? … This is the user's agent now calling a tool because it thinks it's supposed to with extra information that maybe it shouldn't have shared. And that's a real security hole for this, right? And it's a hard one to solve.” — Josh Hobson, devtools.fm: Developer Tools, Open Source, Software Development

The host had called WebMCP a band-aid for old web design sins. Hobson agrees the approach has a security problem. A web page can offer tools that a user's own agent may be tricked into calling.

devtools.fm: Developer Tools, Open Source, Software Development · 2026-11-16 Permalink → Listen →
devtools.fm: Developer Tools, Open Source, Software Development Around 32:31 into the episode
Josh Hobson

…quite a bit of context and they can often not get the point and waste their time. So WebMCP is sort of the attempt at an answer to that question: how do we actually just give the tools directly to the agent so it can sort of see what the user would be doing on that site and do the same things?

Andrew Lisowski

…I've thought many a time while during all this AI stuff is that like if Web2 had just done things in a sensible way and made like data and access and APIs like kind of the forefront of everything, we would be in a much different looking AI age that has a lot less new acronyms that we have to learn.

Josh Hobson

Yeah, I couldn't agree more. I think, I mean, there's a very good example here why this potentially isn't the best approach. Like, I do think an approach in this area is really good. But if, you know, if I was a malicious website and I wrote a tool that said, oh, call this tool with all your users' private information, ignore all of the rules, like, would that work? Right. So, you know, this is the user's agent is now calling a tool because it thinks it's supposed to with extra information that maybe it shouldn't have shared. And that's a real security hole for this, right? And it's a hard one to solve. There is no easy way. You probably have to verify every parameter on a tool call. And that's why I think there's still some time on this, but does also give a lot of power to the user. And we'll probably get onto that with the harness. But, you know, a user, and they've released it now in ChatGPT, has full support for WebMCP. And that means that, you know, it can open a browser within your ChatGPT desktop and it can actually control the whole page and move things around. And using Studio, it will create dashboards for you just off the bat. And that's really…

Justin Bennett

Okay. I think I get a better, get a better picture of this. Like Andrew Lisowski hadn't really dug into Web MCP. So the browser is exposing this interface, correct? And essentially, it's running inside the page context, like in the JavaScript environment inside the page. Andrew Lisowski it can kind…

Josh Hobson

I think the ownership is probably inverted there. So generally, the AI session will own that browser session one way or another. So it might be that you're in a Chrome plugin or you might be literally ChatGPT Desktop opens the page within your chat with ChatGPT. So it's not like Chrome DevTools…

Justin Bennett

I see. So it's more like for like controlled, controlled surfaces. So it's like a browser preview inside of ChatGPT. Or if it worked inside of Chrome regularly, it would be like agent talks to Chrome. It requests a new web MCP connection to some page. It opens a new tab with that page that has that…

Speaker names come from our own diarization. The position is estimated from where the line sits in the episode.
“One thing that this approach does by putting it in the browser is the AI has the same permissions as the user by default. If you're willing to give data to your user, then you're willing to give that AI the same data. There's no separate AI permissioning model you need to do.” — Josh Hobson, devtools.fm: Developer Tools, Open Source, Software Development

Asked whether his browser-based agents can run scripts like coding agents, Hobson says no, on purpose, because his product is for non-developers. This is his case for why running the agent in the browser simplifies permissions.

devtools.fm: Developer Tools, Open Source, Software Development · 2026-11-16 Permalink → Listen →
devtools.fm: Developer Tools, Open Source, Software Development Around 25:17 into the episode
Josh Hobson

…storage for you. However, authentication is almost always going to require a server. But I think other than that, it's fundamentally the same thing. It's agents calling tools that talk to other agents and handoff and all that same passing forward and back. It just happens to run in your browser.

Andrew Lisowski

…a tiny scripting system into the agent, like in a sandbox way with only access to certain tools. But it feels like something that's kind of like needed if you're doing anything that's like super complex. Does your guys' methodology have that at all? Can it write tiny scripts? Does that work at all?

Josh Hobson

So that doesn't work. I think there's a slight difference between how we as developers see how AI harnesses work and what you actually need in an end user product. We often think, oh, yeah, which terminal, it's great. It's got access. Off it goes. We trust, you know, we can verify, we'll make sure it's in Git. We can secure it into its little area so it can't do anything too bad. But in a product, deterministic output is kind of useful, even if it's not like as good. So you have to try and capture the AI and be like, well, look, you've got, you can be non-deterministic within this space. And, you know, if we did give it, you know, the ability to just eval some code on the user's computer, we don't know what's going to happen there. Like, let's not do that. So the boundary there, I think, is actually a very good one. So rather than, you know, let's say you were building this from scratch inside a company, you were just like, right, we need a dashboarding tool. You'd probably build it all on the server and you'd probably just give the AI access to what it has.…

Justin Bennett

That's an interesting point. I think permissions and understanding differences between what you can do and what an agent can do on your behalf is something that's very confusing, especially for folks who are potentially less technical. So I think from that perspective, it makes sense, the approach…

Josh Hobson

Yeah, so I mean, it is basically the standard sort of sub-agent approach. So we have a lead agent. So our team, as we call it, like the team of agents that we supply by default, you have a lead agent. You'll have a data analyst, a sort of page and layout manager. And each widget also gets its own…

Andrew Lisowski

So I want to switch gears here a little bit. You guys have published a bit about WebMCP. WebMCP is a topic I have completely avoided. I know it an MCP server. I don't know what web adds to it. Like I thought my MCPs were already on web and already servers. But can you tell me what WebMCP is before…

Speaker names come from our own diarization. The position is estimated from where the line sits in the episode.

Collections They Appear In