Context
Kath Korevec is a product lead for ChatGPT Sites at OpenAI, working alongside Codex. Her team used Sites internally for over a year and launched it publicly two months before this live DevDay recording. On the day of the episode, OpenAI announced plugins and Sites, so that a site can use the visitor's own connected tools. For PMs, the episode shows how a product lead builds small, personal and team software without engineering help, and what changes when a site can safely read each viewer's own data.
The Big Idea
Sites turns a one-off vibe-coded page into real software: it deploys for you, and each visitor sees their own data through their own connected plugins.
That removes the two hard parts of internal tools, hosting and API keys, so a PM can build a tool for a week, a team or a community and throw it away when it stops being useful.
Key Insights
Plugins let one site serve everyone
- What: a site can now use the visitor's connected plugins, so everyone opens the same site but sees content based on their own account and team.
- Why it matters: there are no API keys to manage. Kath says you "just get that magically connected."
- Example: her incident command site asks each visitor to approve connections to Notion, Slack and calendar. If she shares it with the Codex team or the identity team, they see their own incidents, not hers. This is available for enterprise workspaces.
- Fit: Kath says data teams plugging into a warehouse, and finance teams with sensitive data, like it most, because the visitor's access scopes what the site shows.
A PM's incident view
Kath built an incident command overview because she wants to follow incidents without pinging the team. Her points:
- She spends most of her time in meetings and does not want to get in the way of people solving the incident.
- The site pulls a minute-by-minute timeline of incidents and who is working on each one.
- It reads her approved Slack channels and assumes she is on the Sites team, then shows that team's content.
- It pulls run books and guides from Notion and links straight into them.
- The demo content was scrubbed, since the real data is incident content.
Say "bring your own connector"
Codex reads intent, so a prompt like "build me a site that cross-references Notion documents with my calendar" usually makes it reach for connectors on its own. It may ask if you want plugin connectors, or just build it and explain what it did. Kath's own magic phrase is "bring your own connector". About 60 connectors work today, and her most-used are Slack, Notion, Google Drive and Google Calendar.
Sites is infrastructure, not just pages
- Sites started as the "last mile" for software Codex builds locally: it handles the deploy and DevOps that people struggled with.
- It comes with D1 data, an R2 bucket and co-editors, so it supports full internal tools.
- Users also use it for plugins, hosting data and running evals.
- The announcement that day: you can host MCP plugins through Sites and use them in other sites. Kath calls this a virtuous cycle, since a site can be both the product and the infrastructure behind another one.
Disposable software fits how people think
Kath says her brain works differently from other people's. She might want a different view of Google Calendar for a week, such as when she is on a business trip, so she builds a site for that week and throws it away afterward. She sees users build everything from small prototypes to personal software to deep team tools. The point is that you adapt the tool to how you work, instead of bending to a rigid product.
A skill can distribute a game
- Kath built a large community dungeon crawler on Sites. Players pick a character, move around and fight.
- Instead of an in-app editor, she wrote a "Dungeon Sight" skill that tells Codex the room dimensions, player and controls. Users install it and ask Codex for a room, such as a Stranger Things one, which it built locally on its own.
- The skill asks whether the room should be open-ended, so it can link to another dungeon, or a standalone ending.
- Her plan is for people to share rooms and she adds them to the longest dungeon crawler. Claire Vo called this a new user experience: you download a skill and the agent builds into the app for you.
Speed drives creativity
- Kath's view: the faster models get, the more creativity you get, since waiting on a thread to process breaks her flow.
- Claire tried an early "ultra fast" model announced at DevDay on a 3D moon-ship game with three characters. Typed requests like "Pip needs bunk beds", "it's fall in Momo's garden" or "remove gravity" changed the scene in real time.
- The cost point is real: she said it was about eight times as fast, but as Sam said on stage, about six times as expensive. Her kids kept asking for more changes, and she called it a poor use of her spend.
Bringing inference into a site
Kath's team is testing a widget on a site that lets a viewer ask ChatGPT to change it, or do things from inside the site. The widget was simple and glitchy in the demo. The idea Claire raised: a viewer could bring their own inference, like a kid with an allowance of tokens asking a game to add an octopus. She noted this means choosing carefully what to run in Sites versus what to deploy in your own cloud.
Mental Models & Frameworks
Same site, different data per viewer
A site is a shell, and the plugins decide what fills it. Build one site for a whole org, then let each person's approved connections define their view. Use it for dashboards, team hubs and anything where role-based data matters. Kath's incident site is the example: one URL, different content for each team.
Sites as software home
- Build: Codex writes the code from your prompt and intent.
- Place: Sites handles deploy, so there is a URL to share.
- Connect: plugins bring live data without keys.
- Distribute: a skill can teach an agent how to extend your app.
Use it when you want to move from idea to something others can open the same day.
Trade-offs & Nuance
Light analytics for now
Kath said developers get lightweight analytics: visits and traffic. They cannot see who builds with a skill they published. Her team plans more robust analytics, including investigations and spikes that are more agentic. For now she wants her game to grow through community sharing and tags on Twitter instead of tracking.
Infrastructure limits are real
Claire pointed out that if many people add rooms or bring their own inference, there will be demands on infrastructure. Builders need to decide how much belongs in Sites and what should be deployed in their own cloud with their own infrastructure.
Common Mistakes
Mistake: Letting AI speak in your name
- Why it happens: agents can email or message people on your behalf, and Kath says it makes assumptions.
- Why it hurts: she says signing her name to something it wrote and sending it crosses the line. Claire once had a bot post in her community Slack and people engaged as if it were her.
- Better approach: tell the agent plainly "do not do that". Kath uses it for research and help, and wants her messages to remain her own voice.
Practical Application
Build a team status site
- Do: ask Codex for a site that reads your team's Slack channels and Notion run books, then approve the connectors.
- Then: share it so each teammate sees their own scope.
- Why it works: you follow work without interrupting the people doing it.
Build a throwaway trip site
Before a business trip or a busy week, ask for a site that organizes your calendar the way you need it. Delete it afterward. This takes minutes and removes the pressure to build a permanent tool.
Write a skill for your product
If your product has a creative or building side, write a short skill that explains its structure and controls, and let users' agents extend it. Kath's skill covers room size, the player and the controls.
Automate a weekly discovery task
Kath uses computer use to browse Reddit's top playlists, then runs an automation every Monday at 8 a.m. that builds a new playlist. Pick a recurring research task of your own and schedule it the same way.
Questions to Consider
- Which of your team's internal dashboards currently shows everyone the same data, and what would change if each viewer saw only their own team's data through their own connected tools?
- What tool could you build for one week and then delete, instead of asking engineering for something permanent?
- If users could install a skill that lets their agent extend your product, what would that skill need to tell the agent about your product's structure and limits?
- Where would you draw the line on an AI acting in your name, such as sending emails or Slack messages, and have you told your tools that clearly?
Bottom Line
ChatGPT Sites removes hosting and API keys from internal tools, so a PM can build site software that shows each viewer their own live data. Kath's lesson is to build small, disposable and personal tools, share skills so others can extend them, and keep AI out of speaking for you.
Tools & Products
| Tool / Product | What it does | Why it was mentioned |
|---|---|---|
| ChatGPT Sites | Hosts and deploys software built with Codex, with D1 data, an R2 bucket and co-editors | The subject of the episode, used across OpenAI for slides, incident tracking and games. |
| OpenAI Codex | Coding agent that builds the site from a prompt | Builds the sites and the connector wiring, and reads the intent behind a prompt. |
| Plugins in Sites | Let a site use each visitor's own connected tools | The feature launched at DevDay that makes one site show different data to each viewer. |
| Awesome Sites | Community gallery of sites people built | Kath's gallery at awesomesites.ai, for getting inspired by what others build. |
People to Follow
Kath Korevec
Product lead at OpenAI working on Sites and Codex, who spent over a year building and using Sites internally. She shares her builds and asks for feedback on X as simpsoka.
Notable Quotes
"You don't have to snap to something that's really rigid." (Kath Korevec)
"I want to protect that. That's my voice." (Kath Korevec)
"My brain works in different ways than other people." (Kath Korevec)
