@mcp-layer/schema

@mcp-layer/schema

@mcp-layer/schema extracts MCP server definitions (tools, resources, prompts, templates, and MCP Apps metadata) and normalizes them into a unified Zod-backed schema. The output is designed for downstream generators that need a single contract to build CLIs, REST endpoints, UI renderers, or additional MCP layers without re-parsing the MCP protocol surface.

Table of Contents

• Installation
• Usage
• What this package does
• Output shape (authoritative)
• MCP Apps support
• Generator guidance
• Responsibilities and lifecycle
• Error handling
• JSON Schema vs Zod
• Runtime Error Reference

Installation

pnpm add @mcp-layer/schema
# or
npm install @mcp-layer/schema
# or
yarn add @mcp-layer/schema

Usage

import { load } from '@mcp-layer/config';
import { connect } from '@mcp-layer/connect';
import { extract } from '@mcp-layer/schema';

const config = await load(undefined, process.cwd());
const session = await connect(config, 'demo');
const output = await extract(session);

console.log(output.server.info);
console.log(output.items[0]);
await session.close();

What this package does

• Reads the MCP server capabilities and metadata from the live client connection.
• Calls MCP list endpoints (tools/list, resources/list, resources/templates/list, prompts/list) using pagination.
• Normalizes everything into a unified, type-discriminated schema so generators can consume a single list.
• Wraps tool input/output JSON Schemas into Zod validators backed by Ajv, while preserving the original JSON Schema.
• Converts prompt arguments into a lightweight JSON Schema so prompts and tools share the same detail.input shape.

Output shape (authoritative)

The package returns:

{
  server: {
    info,          // initialize.serverInfo
    capabilities,  // initialize.capabilities
    instructions   // initialize.instructions (optional)
  },
  items: [
    {
      type,         // "tool" | "resource" | "resource-template" | "prompt"
      name,
      title,
      description,
      meta,         // icons, annotations, _meta
      detail        // type-specific data
    }
  ]
}

Type discriminator:

• tool, resource, resource-template, prompt
• MCP Apps do not introduce a new type; tools remain tool and UI resources remain resource.

detail: tools

detail: {
  input: {
    schema, // Zod schema (Ajv-backed)
    json,   // original JSON Schema
    error   // optional Ajv compile error message
  },
  output: { schema, json, error }, // only when the tool advertises outputSchema
  ui: { resourceUri, csp, permissions } // when _meta.ui is present (MCP Apps)
}

detail: resources

detail: {
  uri,
  mimeType,
  size,
  ui: { resourceUri, csp, permissions } // when _meta.ui is present (MCP Apps)
}

detail: resource templates

detail: {
  uriTemplate,
  mimeType
}

detail: prompts

detail: {
  input: {
    schema, // Zod schema derived from prompt arguments
    json,   // JSON Schema derived from prompt arguments
    error   // optional Ajv compile error message
  }
}

MCP Apps support

If a server exposes MCP Apps metadata:

• Tools may include _meta.ui.resourceUri pointing to a ui:// resource.
• UI resources may include _meta.ui with CSP and permissions.

This package normalizes those fields into detail.ui so generators can:

• Detect which tools expect UI rendering.
• Resolve and render the ui:// resource.
• Apply CSP and permissions when hosting the UI in a webview or browser sandbox.

Generator guidance

Suggested uses for the unified schema:

• CLI: Use name, description, and detail.input to build flags and help text.
• REST: Use name for endpoints, detail.input.json for request validation, and detail.output for response schema.
• UI: Use detail.ui.resourceUri to locate the UI and read detail.ui.csp / permissions.

Responsibilities and lifecycle

• This package does not open or close connections. It expects a live Session from @mcp-layer/connect or @mcp-layer/attach.
• You are responsible for calling session.close() after extraction.
• If the server doesn't advertise a capability (tools/resources/prompts), extraction skips that surface.

Error handling

• Schema compilation uses Ajv under the hood. If a tool exposes an invalid JSON Schema, extraction still succeeds but detail.input.error or detail.output.error will be populated for that item.
• Servers that return empty lists produce an empty items array (not an error).
• Servers that support resources but not resource templates are treated as having zero templates.

JSON Schema vs Zod

Zod validators are produced to allow direct runtime validation, but the original JSON Schema is preserved in detail.input.json and detail.output.json so generators can emit OpenAPI, JSON Schema, or other schema formats without losing fidelity.

Runtime Error Reference

This section is written for high-pressure debugging moments. Each entry maps directly to schema extraction preconditions.

Expected a Session instance.

Thrown from: extract

This happens when extract(link) is called without a connected Session object (or with an object that does not expose .client). The extractor relies on MCP client capabilities and list calls.

Step-by-step resolution:

• Verify you are passing the Session returned by connect(...) or attach(...).
• Confirm session setup completed and session.client is available.
• Avoid passing raw MCP client objects directly; wrap flows around Session.
• Add an integration test that extracts from a real connected session.

Fix Example: extract schemas from a connected Session

const session = await connect(config, 'local-dev');
const catalog = await extract(session);
console.log(catalog.items.length);
await session.close();

License

MIT