kata

Inspiration

What it does

KATA

KATA is a system that allows users to teach a web application a new agent capability simply by demonstrating a workflow.

Instead of requiring developers to manually write a dedicated API or WebMCP integration for every repeated action, a user performs the task a small number of times. KATA records the semantic application actions behind those demonstrations, compares the traces, identifies what stays constant and what varies, infers typed parameters, compiles the repeated behavior into a deterministic workflow, and installs it as a new WebMCP tool.

For example, a researcher may repeatedly save a paper, assign it a priority, and add an “architecture” tag. KATA can recognize the common structure and create a tool such as:

triage_architecture_paper(workId)

An AI agent can then discover and execute that newly learned capability through WebMCP.

The important distinction is that KATA does not record mouse coordinates or replay brittle DOM clicks. It records semantic application commands. The human UI and the agent both ultimately use the same command system, so a learned tool invokes real application behavior rather than a separate simulation.

KATA also avoids generating arbitrary executable code. Learned workflows are represented as structured, validated programs built from a finite set of supported commands. Inputs are validated before execution, the complete batch is preflighted before mutation, and invalid workflows fail without partially applying actions.

The deeper idea behind KATA is user-extensible application APIs: the API available to an AI agent no longer has to be completely defined by the developer ahead of time. Users can teach software the workflows that matter to them, and the application can safely turn those workflows into agent-usable capabilities.

How we built it

Challenges we ran into

Accomplishments that we're proud of

What we learned

What's next for kata

Built With

• deterministic-workflow-compiler
• html/css
• javascript
• localstorage
• openalex-api
• semantic-command-bus
• serverless
• structural-anti-unification
• typed-json-programs
• vercel
• webmcp