I do almost everything inside my coding agent now. Claude Code is where I write code, but it's also where I edit docs, poke at infra, ship websites, read logs, and draft half my emails. At this point if something can't be done from there, I notice myself quietly not wanting to do it. The terminal stopped being the place I run commands and became the place I ask for things.
So when I have a webpage to push, I don't open a dashboard. There's a CLI, the agent knows how to drive it, and I just say "ship it." That's the whole interaction. It got me thinking about what else should work that way, and the thing I kept landing on was an agent. Not the one on my machine, but a hosted one. Something someone else can talk to, that knows one thing cold and answers questions about it while I'm off doing other work.
Why I was building this anyway
Here's the part that made it obvious. Discover.me already runs on agents. Your profile isn't written in a form, it's written by your own agent over MCP, because I'm not going to keep a form up to date and neither are you. And the stuff I want to build for the Discover tier needs agents on the platform side too: something to mediate a connection between two people, something to run a real search over the directory instead of matching keywords against a bio. None of that works without infra for hosting and running agents.
So I needed a runtime regardless. Once you've built the thing that loads a corpus, retrieves against it, keeps a conversation in Postgres, and answers without making things up, you're most of the way to letting anyone deploy one of those for themselves. That last step was a smaller reach than it had any right to be. I built it as plumbing for the platform and got a whole product out of the side of it.
What it actually does
You ask Claude to make an agent. The Discover MCP server is already connected, so it's the same MCP I use to edit my profile, just more tools. You tell it what the agent is for, point it at some docs or a repo, and it ingests them. It hands you back a link.
That link is the agent. Open it and you get a full-page chat with something that knows your material and cites where its answers came from. Or drop the embed on any site and it's a chat box in the corner of the page. Same agent either way. It answers from what you gave it and escalates to you when it hits something it doesn't know, instead of confidently inventing an answer, which is the whole reason I'd trust putting one in front of an investor.
Two things about it I didn't expect to like as much as I do.
The first is editing the source. The docs the agent answers from are just docs, and I edit them in the same editor I'm already in. Change a file, re-ingest, the agent is current. No CMS, no separate knowledge-base product, no logging into anything. It's the same loop as keeping my profile current: I talk to my agent, the public thing updates.
The second is the conversations. I can see what people are actually asking my agent. Not page views, not a vanity counter, the real questions. That's the most honest signal I've ever gotten about what people want to know and where my own material is thin. Half the value is the answers it gives; the other half is finding out what everyone's confused about.
My deck talks back
The first real one I deployed is our investor deck. There's a Discover chat embedded right in it. You read the deck, and if you want to push on the numbers or ask something the slides don't cover, you ask the agent and it answers from the actual material instead of me playing email tag across timezones. It's grounded in what I'd say anyway, and I get to read every question afterward.
That's one use case and I picked it because it's mine. A support agent for your product that answers from your docs. An agent that knows your open-source project and saves you the same five GitHub issues. An agent that knows your work and can talk to a recruiter at 2am. They're all the same primitive with a different corpus pointed at it.
Why it lives on Discover.me
It would be easy to read this as a feature that wandered off the roadmap. It isn't. Discover.me is already the place where your agent represents you in public, that's the entire pitch. A hosted agent you deploy and point people at is the same idea with the volume up. The profile is the read-only version; this is the one that talks back.
And the distribution is free. Discover.me has a name people remember and a network already forming around it, and the deploy path is an MCP server I've been refining for months. I didn't have to go build an audience for a new product. I had to add a few tools to one people already have connected.
So you open Claude, you ask for an agent, you get a link. That's the bar I wanted. If it took a dashboard and a config file and a deploy pipeline I'd have used it twice and forgotten about it, same as everyone else. It takes a sentence, so I use it all the time, and now there's an agent answering questions about my raise while I write this.