Who actually shows up to a Claude night in San Francisco

Between September 19 and September 30, the Claude community in San Francisco ran four events: two Build Days, a spatial computing workshop and a Physical AI night.

The registration forms asked a few short questions. What is your experience level with Claude Code? What do you do? Why are you coming? Which Claude Code features do you want to discuss?

Not every event asked every question, so each percentage below comes from the people who answered that one. I organize these events, and I went in with assumptions. Some held up. Most didn't.

Almost everyone is already a daily user

79% said they use Claude Code every day. Add the people who use it several times a week and you get 93%. Almost nobody described themselves as new.

That changes what a good event looks like. Nobody in that room needs an introduction to Claude Code. They want the thing they haven't tried yet: hooks, subagents, the Agent SDK, connecting Claude to real systems. The intro should be short, and the build block should be long.

More founders than engineers

Of the people who told us their role:

• Founders and executives: 35%
• Engineers: 29%
• Students: 15%
• GTM, ops and other: 10%
• Product and design: 6%
• Researchers: 6%

And the founders are early: 84% of those who shared a stage were at idea or pre-seed.

Company size tells a similar story: 62% came from companies of up to 500 people, but almost one in five came from companies with more than 5,000.

These are people deciding what to build next, many of them without an engineering team yet. When they build at a Claude event, it isn't a toy. It's a first prototype.

People come for each other

We asked why people were coming, with more than one choice allowed:

• Meet collaborators or cofounders: 70%
• Learn what's possible: 49%
• Get feedback on what they're building: 35%
• Find hardware or demo collaborators: 34%
• Meet investors: 25%
• Find a role: 22%
• Hire: 7%

Meeting people beat learning. I'd been designing these nights around content: talks, demos, a good deck. The data says the content is the excuse. The room is the product.

One more number worth sitting with: three times more people came looking for a role than came to hire.

What they want to talk about

Robotics, hardware, 3D and spatial topics led, which makes sense: two of the four events were about spatial computing and physical AI.

Outside hardware, the list is a snapshot of where builders are right now, in order:

1. Subagents and multi-agent orchestration
2. Skills, hooks and plugins
3. MCP and tool integrations
4. New models (Opus 5.5 and Fable 5.1)
5. Computer use, vision and multimodal
6. Long-running and autonomous loops
7. Safety, guardrails and governance

Multi-agent work leads, and the plumbing that makes agents reliable (hooks, skills, MCP) is right behind it. Interest in the new models is real, but builders seem more curious about how to put agents together than about which model runs them.

The room wants to give back

The last question was how people wanted to get involved. Dozens asked to demo or speak at a future event, dozens more to help organize, and ten offered their company's space as a venue. People from 30 universities registered, led by UC Berkeley, Stanford and San José State.

That's the most encouraging part of the whole set. A community isn't an audience. Four events in, the people in the room were already asking to help run the next one.

What I changed

1. Shorter intros, longer build blocks. Daily users don't need the basics.
2. Design for founders too: show how far you can get without an engineering team.
3. Treat meeting people as the main event: team formation, one-minute demos, real time to talk.
4. Give the topics people asked for their own slot: multi-agent patterns, hooks and skills, MCP.
5. Say yes to the people who want to demo and to the companies that offered a venue.

How these numbers were collected

Optional questions on the Luma registration forms of four events, September 19 to 30, 2026, counting approved guests only. Not every event asked every question: each percentage comes from the people who answered that one, between about 150 and 550 answers. These are registrations, not attendance. No individual answers are shared.