mcp-discovery

Automatically discovers and registers MCP (Model Context Protocol) servers running on the local machine by scanning standard configuration directories and environment variables, then dynamically loads their tool schemas without requiring manual server URL configuration. Uses filesystem introspection and MCP protocol handshakes to build a registry of available tools at runtime.

Extracts and validates tool schemas from discovered MCP servers by parsing their protocol responses, normalizing schema formats across different server implementations, and validating tool definitions against MCP schema standards. Builds a unified tool registry that abstracts away server-specific schema variations.

Routes discovered tools to an LLM (via OpenAI, Anthropic, or other compatible APIs) using function-calling protocols, allowing the LLM to select and invoke appropriate tools based on user intent. Handles parameter binding, error handling, and result formatting to integrate tool outputs back into the LLM conversation context.

Manages the lifecycle of discovered MCP servers including connection establishment, health monitoring, graceful shutdown, and error recovery. Maintains persistent connections to servers and handles reconnection logic if servers become unavailable, ensuring reliable tool availability throughout the LLM agent's execution.

Abstracts LLM provider differences by supporting function-calling APIs from OpenAI, Anthropic, and other compatible providers through a unified interface. Translates tool schemas and function-calling requests/responses between provider-specific formats, allowing the same agent code to work with different LLM backends.

Maintains execution context across tool invocations including conversation history, tool call results, and agent state. Provides a stateful execution environment where the LLM can reference previous tool outputs and the agent can track which tools have been called and their outcomes, enabling multi-step reasoning and tool chains.

Implements structured error handling for tool invocation failures including timeout management, parameter validation errors, and server-side tool errors. Captures error details and passes them to the LLM for recovery decision-making, allowing the agent to retry failed tools, try alternative tools, or gracefully degrade functionality.

Generates human-readable documentation for discovered tools including descriptions, parameter requirements, return types, and usage examples. Provides introspection APIs that allow developers to query tool capabilities, list available tools, and inspect tool schemas at runtime for debugging and UI generation.