Systems & Research
The Rectangle
A working creative space in Santa Cruz for DJ rehearsals, media production, and XR experiments, supported by an agent-ready booking website.
Agent-Ready Web Applications
Web apps for people and AI agents.
When an AI agent uses a website it clicks through the interface or scrapes it. That breaks when the layout changes, and the business cannot tell a harmless lookup from an action that sends money or personal details.
For businesses whose website takes inquiries, bookings, or orders and who want AI agents to use it correctly, with a person in control of anything that has consequences.
Worked example
An app a person and an agent use together
WebMCP Replay Room is a tennis replay workspace where a person and a browser agent review shots together. Six capabilities are defined once with Agent Forge, three read and three reversible-write, and registered as tools in the browser over WebMCP. Without WebMCP the app runs as an ordinary interactive demo.
A person selects a shot that feels wrong. The agent reads it, finds similar ones, puts them in the review rail, and compares them with stronger references. On request it writes the issue into the coaching plan. Every change is visible and reversible, and none of it clicks or scrapes the interface.
The same layer runs The Rectangle’s website and this one: each page registers its tools, consequential actions are flagged so an agent confirms with the user first, and Markdown alternates and llms.txt give agents the same facts people read.
How it works
- Agent Forge
defines each operation once with a schema and an effect class and publishes it to browser agents through WebMCP.
- Agent-native web apps
are the server-rendered Nuxt apps the tools live in, fast without JavaScript and readable by crawlers and agents.
- Agent City
turns the execution events the tools emit into a view of what agents did, without logging what users typed.
What we can build
Typed tools
Each operation defined once with a schema and an effect class, published to browser agents through WebMCP.
Approval flags
Actions that send money or personal details marked consequential, so an agent confirms with the user first.
Server-rendered apps
TypeScript, Vue, and Nuxt apps backed by Postgres, fast without JavaScript and readable by crawlers and agents.
Discovery surfaces
Markdown alternates, llms.txt, and structured data, so agents read the same facts as people.
Inquiry pipelines
AI screening, routing, CRM sync, and a person’s approval before anything is invoiced.
Observability
Metadata-only execution events, so you can see what agents did without logging what users typed.
Projects
Projects in this area.
Systems & Research
WebMCP Replay Room
A tennis replay workspace where a person and a browser agent review shots together through six reversible Agent Forge capabilities.
Systems & Research
Agent Forge
An open-source TypeScript library that lets a web app expose its real operations to AI agents as typed tools, each with a declared effect and a flag for consequential actions.
Where to start
One action, published to agents
We take one existing form or action on your site, define it as a typed tool with the right effect class, and add Markdown alternates for the pages agents read most.
- What you get
- The tool live on the page, the Markdown alternates and llms.txt, and a note on which other actions are worth publishing and which need approval.
- What you bring
- The site and the action to start with
- Who approves a consequential action
- Access to the repository or a staging environment
Next step
Make your site agent-readyGet in touch
Tell us what you have in mind.
Describe it in a few sentences. You get a draft brief to review before anything is sent.