MCP Night 2025: When the AI infra community overflowed the Exploratorium in San Francisco

On May 14, 2025, we threw the first-ever MCP Night—a free evening event at the Exploratorium in San Francisco dedicated to one of the most exciting developments in applied AI: the Model Context Protocol (MCP).

What started as a modest idea to get a few MCP builders in the same room turned into a full-blown community convergence.

We had hundreds of engineers, founders, researchers, and tinkerers pack out the venue—so many, in fact, that we had to spin up a whole second demo area just to make room for all the action.

MCP: The Interface between models and the real world

If you’re new to MCP, here’s the one-sentence pitch: it’s how you connect large language models to your actual tools, APIs, data, and systems—safely, securely, and in a standardized way.

MCP enables a model to generate text and perform actions such as filing tickets, retrieving customer data, calling APIs, managing files, and automating real workflows.

It defines a protocol for exposing capability, access control, and input/output semantics in a way that lets LLMs reason about your software environment.It’s a big deal. And it’s growing fast.

WorkOS AuthKit secures your MCP servers and tools, allowing you to configure granular access based on roles.

AI systems improve as they are given access to more data, and Authkit helps you do this securely.

Who demoed what

We invited teams building some of the most exciting MCP Servers and integrations to demo what they’ve been working on.

Main stage presenters included:

Kylie Czajkowski of Cloudflare, showing how you can wire LLMs into network and edge automation with Cloudflare workers.

Paul Klein of Browserbase showed how the browser may be the only MCP server you need.

Angie Jones of Block demonstrated Goose, an internal agent turned open-source personal assistant for developer productivity.

Mattt Zmuda of Loopwork showed how MCP can help you find contacts and customers using natural language.

David Cramer of Sentry, showed how MCP and Sentry will enable workflows that can auto-remediate issues by opening pull requests to fix bugs detected in production systems.

‍Dustin Schau of Postman demonstrated how to use public APIs to create a Node.js boilerplate that wraps the endpoints to auto-create tools and MCP servers for you.

‍Andre Landgraf of Neon, showed how MCP is the ideal protocol for creating excellent developer experiences when building full-stack applications.

‍Andrew Qu of Vercel, who showed how you can use MCP via the AI SDK to ask questions about your Vercel deployments and environments via Cursor.

Chris Bell of Knock showed us how to build intelligent notification systems that can be queried and triggered by LLMs.

Cameron Matheson of WorkOS showed off how to wrap MCP servers with Authkit to expose auth, audit logs, and org metadata via MCP.

The non-deterministic nature of LLMs always adds an extra level of intensity to a live demo!

Overflow demos and a museum-wide afterparty

The turnout far exceeded expectations.

The main stage quickly reached capacity, and we expanded into an overflow demo area to accommodate the number of teams eager to showcase what they’d built.

Once the talks concluded, the Exploratorium opened up for the afterparty.

With exhibits active, music playing, and a well-stocked bar, the atmosphere shifted from technical deep dive to casual, high-bandwidth exchange.

Conversations about AI, infrastructure, and auth continued across science installations, over drinks, and into the night.

And yes, Mr. GPT, a local favorite and unofficial mascot of San Francisco’s robotics scene, made an appearance, reminding us that while MCP stands for Model Context Protocol, that night it stood for music, cocktails, and partying!