I built a prototype showcasing how I think AI assistants will interact with existing software during an OpenAI Realtime Hackathon this weekend. Won 2nd place and $15k in credits. (skip to 1:00)

I built a prototype showcasing how I think AI assistants will interact with existing software during an OpenAI Realtime Hackathon this weekend. Won 2nd place and $15k in credits. (skip to 1:00)

https://hirebellhop.com

Circumnavigated earth by train and ship while building tech companies
|| 
New York <-> Tokyo
|| 
reustle.org

Pretty slick.

Think this could work without a formal integration? As in, login to a saas it’s never seen, invoke it and have it discover what it can do.

memebuilding • prev manifold.xyz, twilio.com

There are a few routes to go down w/ that.

What I've been brainstorming for a few weeks, heavily related to MCP, is how to essentially let these agents go hunt down the API docs for the service & your API key when logged in, then it can do most (?) things w/out fumbling w/ the browser.

That's one way I was thinking, you give it a week or something and it monitors, but feels fragile and doesn't adapt to changes easily.

I'm going to spend more time on this

That's one way I was thinking, you give it a week or something and it monitors, but feels fragile and doesn't adapt to changes easily.

I'm going to spend more time on this https://github.com/orgs/modelcontextprotocol/discussions/84

Yeah standards like agent.txt could help.

I wonder if there’s just a training phase where the extension traces network calls while you do basic tasks.

If you export a HAR file in chrome for example it has all the network details and the stack traces. That plus knowledge of the DOM might be a good start.

Extension could even say when it’s discovered a skill from watching you use the browser.

yeaah have been wondering about this the other day