nanoMCP

中文 | English

"The question is not what you look at, but what you see." — Henry David Thoreau

Minimal MCP (Model Context Protocol) demos in Python:

• a tiny MCP server that exposes a few toy tools (add, multiply, weather)
• MCP clients for stdio, SSE, and Streamable HTTP
• an OpenAI function-calling loop that calls MCP tools
• Tavily MCP integration examples (local stdio via npx, and remote Streamable HTTP via SSE)

This repository is intentionally small and easy to read: each transport is a self-contained end-to-end example.

If you forget what an “agent loop” is or how OpenAI function calling works, read: https://github.com/sanbuphy/nanoAgent

Version

v0.1.0

Docs

There is also a docs/ folder with MCP reading notes and deeper implementation analyses:

• docs/README.en.md
• docs/README.md (Chinese)

Install

pip install -r requirements.txt

Environment variables

These examples call an LLM via the OpenAI Python SDK.

macOS/Linux:

export OPENAI_API_KEY='your-key-here'
export OPENAI_BASE_URL='https://api.openai.com/v1'  # optional
export OPENAI_MODEL='gpt-4o-mini'  # optional

Windows (PowerShell):

$env:OPENAI_API_KEY='your-key-here'
$env:OPENAI_BASE_URL='https://api.openai.com/v1'  # optional
$env:OPENAI_MODEL='gpt-4o-mini'  # optional

Tavily examples also require:

export TAVILY_API_KEY='your-tavily-key'

Some scripts support StepFun-compatible env vars (as a fallback):

• STEP_API_KEY (used when OPENAI_API_KEY is missing)
• DEFAULT_BASE_URL (defaults to https://api.stepfun.com/v1)
• DEFAULT_MODEL (defaults to step-3.5-flash)

Quick start (local demo MCP server)

If you're new to MCP, start with stdio/ first. It is the simplest transport and the easiest to debug:

• Read stdio/mcp_stdio_server.py to see the minimal initialize / tools/list / tools/call implementation.
• Then read stdio/mcp_stdio_client.py to see how the client spawns the server, converts MCP tools to OpenAI tools, and runs the loop.

stdio

python stdio/mcp_stdio_client.py "What is 3 + 5?"

SSE (HTTP POST + SSE response)

python sse/mcp_sse_client.py "What's the weather in Beijing?"

Streamable HTTP (HTTP JSON request/response)

python streamable_http/mcp_streamable_http_client.py "Compute 7 * 6, then add 5."

Each client spins up its matching server process automatically.

Tavily MCP examples

1) Tavily over stdio (local npx process)

Requires Node.js so npx is available.

export TAVILY_API_KEY='your-tavily-key'
python stdio_tavily/mcp_stdio_tavily_client.py "Use Tavily search 2 times and give me 5 links about: MCP Streamable HTTP transport"

2) Tavily over Streamable HTTP (remote SSE URL)

This example connects to a remote MCP server that uses the Streamable HTTP pattern (SSE stream + POST message endpoint).

export TAVILY_REMOTE_SSE_URL='https://.../sse'  # optional, a default is baked into the script
python streamable_http_tavily/mcp_streamable_http_tavily_client.py "Use Tavily search 3 times and give me 3 links about: MCP tools/list"

What is implemented

These examples implement the minimal practical subset of MCP tools:

• JSON-RPC 2.0 envelope
• lifecycle handshake: initialize
• tool discovery: tools/list
• tool invocation: tools/call
• (remote Streamable HTTP client only) notifications/initialized

Protocol version used in demos: 2024-11-05.

Project layout

• stdio/
  • mcp_stdio_server.py: MCP server over stdio
  • mcp_stdio_client.py: MCP client + LLM loop (spawns the server)
• sse/
  • mcp_sse_server.py: MCP server over HTTP that replies via SSE
  • mcp_sse_client.py: MCP client + LLM loop (spawns the server)
• streamable_http/
  • mcp_streamable_http_server.py: MCP server over HTTP JSON
  • mcp_streamable_http_client.py: MCP client + LLM loop (spawns the server)
• stdio_tavily/
  • mcp_stdio_tavily_client.py: connect to tavily-mcp via stdio (npx)
• streamable_http_tavily/
  • mcp_streamable_http_tavily_client.py: connect to a remote Streamable HTTP (SSE) endpoint
• docs/: MCP protocol notes + implementation deep dives

Notes

• Default ports: SSE server uses 127.0.0.1:8765, Streamable HTTP server uses 127.0.0.1:8766.
• If the model keeps calling tools, the clients stop after a maximum number of iterations and return Max iterations reached.

mcp references

• MCP architecture overview: https://modelcontextprotocol.io/docs/learn/architecture
• MCP specification overview: https://modelcontextprotocol.io/specification/2025-11-25
• MCP tools section: https://modelcontextprotocol.io/specification/2025-11-25/server/tools
• MCP transports (stdio / streamable http): https://modelcontextprotocol.io/specification/2025-06-18/basic/transports

---

license

MIT

────────────────────────────────────────

⏺ Like a single seed that grows into a forest, one file becomes infinite possibilities.