guzl-aware

A WebMCP reference implementation exploring a simple idea:

WebMCP can tell an agent what it can do. An experience contract can tell it what matters while doing it.

Challenge experiment

This repository is a new, public WebMCP experiment created for the 2026 WebMCP Challenge. It demonstrates a narrow experience-contract pattern inspired by prior LumynQ work without containing or depending on the private LumynQ implementation.

Prior work vs. challenge work

Predates the challenge: LumynQ as a broader experience-intelligence concept/framework and related developer work.

Created in this repository during the challenge: the WebMCP-facing experience-contract schema, tool surface, fictional youth-program demo, visible human-agent interaction, and evals/tests.

Demo thesis

A parent wants to cancel their child's upcoming outdoor camp session. A normal transactional surface can expose cancel_booking. guzl-aware additionally exposes structured experience context so an agent can discover:

• the experience intent;
• what is known and still unknown;
• what it must not infer;
• available alternatives;
• decision/confirmation boundaries; and
• when human support is appropriate.

The contract does not script the agent's response or infer the human's emotions. It provides boundaries and context while leaving the agent responsible for helping the human.

Current WebMCP tools

• get_experience_context — read-only experience intent, knowns, unknowns, non-inference boundaries, and alternatives.
• get_decision_context — read-only confirmation and consequence boundary for cancellation.
• clarify_uncertainty — read-only interpretation of context explicitly supplied by the human, without sentiment-based inference.
• get_human_support_options — read-only support routes when explicit context crosses a defined support boundary.
• report_experience_friction — state-changing structured demo feedback.
• cancel_booking — consequential state change that fails safely unless confirmed: true.

Tool calls are reflected in the visible page activity stream so the human can see what the agent discovered and when state changes.

Run locally

npm install
npm run dev

For local WebMCP testing in supported Chrome builds, enable chrome://flags/#enable-webmcp-testing, relaunch Chrome, then use a WebMCP-aware agent or the Model Context Tool Inspector.

Frozen competition scope

One fictional youth-program scenario. One consequential cancellation action. Five experience-context tools. One visible human-agent loop. Evals. Deployed demo.

Anything that does not materially strengthen that end-to-end demonstration is out of scope for the challenge build.

Status

Working application scaffold and WebMCP tool surface are implemented. Next: deterministic evals, browser verification, deployment, and demo capture.

License

MIT. See LICENSE.