AgentR Universal MCP SDK

Provides a Python-native decorator-based framework for building Model Context Protocol servers without boilerplate. Uses Python decorators (@mcp_tool, @mcp_resource) to register server capabilities, automatically handling protocol serialization, message routing, and lifecycle management. Abstracts away low-level MCP protocol details while maintaining full protocol compliance.

Provides a built-in credential store and injection system that securely manages API keys, tokens, and secrets for MCP servers without requiring external secret management infrastructure. Uses environment variable detection, credential caching, and optional encryption to inject secrets into tool execution contexts. Integrates with common auth patterns (OAuth, API keys, bearer tokens) and supports credential scoping per tool or resource.

Provides testing utilities including a mock LLM client for unit testing MCP servers without external dependencies. Includes fixtures for tool invocation, assertion helpers for validating tool behavior, and support for mocking external API calls. Enables fast, deterministic testing of MCP server logic without network calls or real LLM API usage.

Automatically generates API documentation (Markdown, HTML, OpenAPI) from MCP tool definitions, resource descriptions, and docstrings. Includes tool signatures, parameter descriptions, example usage, and error documentation. Supports custom documentation templates and integration with documentation platforms (ReadTheDocs, GitHub Pages).

Provides abstraction layer for connecting MCP servers to multiple LLM providers (OpenAI, Anthropic, local Ollama, custom endpoints) through a unified client interface. Handles provider-specific protocol differences (function calling schemas, message formats, streaming behavior) transparently, allowing the same MCP server to work with any supported LLM without code changes. Includes automatic schema translation and response normalization.

Automatically generates MCP-compliant tool schemas from Python function signatures and type hints (Pydantic models, native types). Validates input arguments against schemas at runtime, providing type safety and automatic OpenAPI/JSON Schema generation. Supports complex nested types, optional parameters, and default values with minimal boilerplate.

Enables declarative definition of MCP resources (documents, files, data) and prompts (system instructions, few-shot examples) with support for dynamic content generation. Resources can be static files, generated on-demand, or streamed from external sources. Prompts support templating and variable substitution, allowing LLMs to access contextual information without embedding it in every request.

Handles MCP server startup, shutdown, and resource cleanup through lifecycle hooks (on_startup, on_shutdown). Manages connection pooling, credential caching, and external resource cleanup automatically. Supports graceful shutdown with timeout-based force termination, ensuring no in-flight requests are lost and all resources are properly released.

Enables async Python functions as MCP tools, allowing non-blocking I/O operations (API calls, database queries, file operations) without blocking the server. Automatically handles async context management, concurrent tool execution, and error propagation. Supports both sync and async tools in the same server with transparent execution model.

Provides structured error handling that converts Python exceptions into MCP-compliant error responses with meaningful messages for LLM clients. Supports custom error types, error context preservation, and automatic error logging. Distinguishes between recoverable errors (retry-able) and fatal errors (non-retry-able) to guide LLM behavior.

Provides centralized configuration management for MCP servers with environment variable detection, type coercion, and validation. Supports configuration files (YAML, JSON, .env) and environment variable overrides, with automatic fallback to defaults. Enables easy deployment across different environments (dev, staging, prod) without code changes.

Integrates structured logging with support for multiple log levels, formatters, and handlers. Provides built-in metrics collection (request count, latency, error rate) and integration points for external observability platforms (Datadog, New Relic, Prometheus). Logs all tool invocations with context (tool name, arguments, execution time, result) for debugging and monitoring.