Typed MCP tools

Every tool is typed, validated, and ready for an SDK.

JSON Schema in, TypeScript out. The agent gets runtime validation, your team gets autocompletion.

Every generated tool ships with a named input schema and a strongly-typed execute() signature. The same schema powers agent discovery, runtime validation, test fixtures, and copy-paste TypeScript for your backend.

JSON Schema, not prose

Claude discovers each tool with a strict `inputSchema` — enum unions, required fields, format hints — so malformed calls fail fast instead of reaching your DOM.

Generated .d.ts files

The Dev tab exposes a `.d.ts` per tool: `{ToolName}Input`, `{ToolName}Result`, `{ToolName}Tool`. Drop into your agent app for typed calls.

Test skeletons out of the box

Every tool gets a Vitest `test.spec.ts` covering happy path, missing-required rejection, and result shape — so CI catches regressions before agents do.

Generated Zod/JSON schema

inputSchema: {
  type: "object",
  properties: {
    projectName: { type: "string", minLength: 1 },
    template: {
      type: "string",
      enum: ["blank", "react", "node"]
    },
  },
  required: ["projectName"]
}

Other features

Auto intent mapping

Infer what visitors want to do

We crawl every page and turn user intents into typed tools — without you having to name them, describe them, or maintain them.

Live debugger

Run your tools against your real site

Playground-style invocations with the real DOM, real selectors, and a response inspector that speaks MCP.

Agent analytics

See which agents call which tool, and when

Per-tool telemetry wired through Optimizely Data Platform. Know who's using your WebMCP before you over-invest.