Overview
A zero-dependency Python micro-framework for the Model Context Protocol. Sync and async implementations (umcp.py and aioumcp.py) with a shared umcp_shared.py module, no third-party packages, stdio, Streamable HTTP, legacy HTTP+SSE and TCP transports, and a naming convention — functions starting with tool_ or prompt_ are discovered and exposed automatically. Type hints become JSON Schema; docstrings become descriptions. Adding MCP to an existing script takes minutes.
How it works
The protocol layer reads JSON-RPC 2.0 over stdio, Streamable HTTP, legacy HTTP+SSE or raw TCP. It discovers tool_ and prompt_ methods on your subclass via introspection, generates inputSchema from type hints (including Literal, Union, Optional, TypedDict), and infers MCP annotations (readOnlyHint, destructiveHint) from naming conventions. Return a value and it serialises automatically. Pick MCPServer for local-disk and compute work; pick AsyncMCPServer when tools are network-bound.
Features
Zero dependencies
Pure Python 3.10+ standard library; no third-party packages.
Local and network transports
stdio, Streamable HTTP, legacy HTTP+SSE and raw TCP.
Convention over decoration
Name a method tool_ or prompt_ — discovery, schema generation, and MCP wiring happen automatically.
Sync + async
MCPServer for compute / disk-bound tools; AsyncMCPServer for network-bound tools. Same API surface.
Full schema inference
Literal, Union, Optional, TypedDict -> JSON Schema. Strict argument validation, type coercion for stringy clients.
Prompt templates
prompt_* methods exposed via prompts/list and prompts/get. Reusable structured interactions for AI assistants.
Architecture
Posts
2026-04-30
Lessons on Building MCP Servers
2026-01-14
Notes on SKILL.md vs MCP
2025-10-04
Creating Per-Project MCP Servers