Read the semantics, not the scripts.
Server-side intake sanitizes public HTML. Forms, repeated content and tables become candidates, while host policies and redirect checks constrain what can be fetched.
Turn semantic HTML into reviewed WebMCP tools. Inspect the evidence behind each contract, export an adapter, and check the tools a deployed site exposes.
Graft compiles the owned Signal Cabinet fixture into seven candidates: five ready, one held and one rejected. The source preview stays beside the contracts it produced.
Browser capture · owned demonstration catalogA page can be easy for a person to use and still be a brittle interface for an agent. Generating tool names is easy; deciding what a tool means, whether it is safe, and whether the deployed contract still matches is harder.
I built a workbench that derives typed tool candidates from semantic HTML, shows the evidence behind each candidate, and lets an owner review what gets exported.
The engineering behind the screens, in three stages.
Server-side intake sanitizes public HTML. Forms, repeated content and tables become candidates, while host policies and redirect checks constrain what can be fetched.
Confidence comes with reasons. Ambiguous candidates are held for review. An owner can repair names and descriptions before a contract ships.
The export bundles reviewed contracts and their runtime. A separate verifier inspects a deployed site and compares registered tool names with the expected set to expose drift.
An accessible name and a guessed class name are different evidence. The inspector makes that difference visible instead of hiding it behind one score.
A derived write does not automatically become an executable handler. Review is a real shipping boundary.
An independently deployed owner example and OmniDev’s Agent Lab exercise the path beyond the workbench itself.
Interactive workbench, owner-site example and exported adapters.
Generated contracts are review artifacts, not a blanket guarantee of production safety. The public owner example uses a demonstration catalog. Historical verifier results in the repository describe their recorded runs.
Agent-native design starts with an inspectable contract, and continues after deployment.