openmcp-core

Converts OpenAPI 3.0/3.1 specifications into Model Context Protocol tool definitions while preserving JSON Schema type information, parameter constraints, and response structures. Uses a schema mapping layer that translates OpenAPI components (paths, parameters, requestBody, responses) into MCP ToolDefinition objects with full type fidelity, enabling LLMs to invoke external APIs with structured, validated inputs and outputs.

Exports a comprehensive TypeScript type hierarchy for MCP artifacts (ToolDefinition, ResourceDefinition, PromptDefinition, CallToolRequest, etc.) with built-in validation logic that enforces MCP protocol constraints at compile-time and runtime. Uses discriminated unions and branded types to ensure only valid MCP messages can be constructed, preventing malformed tool calls or resource definitions from reaching LLM execution contexts.

Abstracts tool calling across different LLM providers (OpenAI, Anthropic, Ollama, local models) by normalizing their function-calling APIs into a unified MCP-compatible interface. Handles provider-specific quirks (OpenAI's tool_choice parameter, Anthropic's tool_use content blocks, Ollama's function calling format) transparently, allowing developers to write tool-calling logic once and execute against any provider without conditional branching.

Generates MCP ResourceDefinition objects from TypeScript interfaces, JSON Schema, or database schemas, enabling LLMs to discover and access structured data sources (databases, file systems, APIs) through a standardized resource protocol. Maps schema properties to resource templates with URI patterns, MIME types, and access metadata, allowing Claude to query resources with type-safe parameters and receive validated responses.

Provides a system for defining reusable MCP PromptDefinition objects with parameterized templates that support variable substitution, conditional blocks, and composition. Enables developers to create prompt libraries that Claude can invoke dynamically, with arguments bound at runtime, supporting use cases like dynamic few-shot examples, context-aware instructions, and multi-step reasoning templates.

Provides base classes and routing utilities for building MCP servers that handle incoming tool calls, resource requests, and prompt invocations. Implements request/response marshaling, error handling, and protocol compliance checking, allowing developers to focus on business logic rather than MCP protocol details. Supports both synchronous and asynchronous handlers with automatic type coercion and validation.

Handles formatting of tool execution results into MCP-compliant responses, with support for streaming large results, binary data, and error propagation. Automatically converts tool output (strings, objects, buffers) into MCP TextContent, ImageContent, or ResourceContent blocks, and manages streaming responses for long-running operations without buffering entire results in memory.

Validates incoming tool call arguments against MCP ToolDefinition schemas before execution, using JSON Schema validation with detailed error reporting. Automatically coerces argument types (string to number, object to typed class) and enforces required parameters, enum constraints, and range limits, preventing invalid arguments from reaching tool handlers and providing LLMs with clear error feedback for retry.