我这次决定用 Python asyncio 从协议层往上写一遍，把 JSON-RPC 消息编解码、能力协商、工具调用链路全部拆开，最后再通过 TaoToken 统一 Key 接入真实模型服务，跑通一个可扩展的 MCP 完整实现。

MCP（Model Context Protocol）本质上是一套基于 JSON-RPC 2.0 的通信规范，它规定了 Client 和 Server 之间「怎么打招呼、怎么报能力、怎么调工具、怎么回结果」。它不关心你用什么语言写 Server，也不关心对面是 Claude、Cline 还是你自己写的脚本。只要消息符合规范，就能通。

Hello, this is Takumi.The history of AI coding is currently being rewritten.The speed of evolution for 'autonomous AI agents' like Claude Code, Cursor, and Google's Antigravity is staggering. The ...

Introduction: The Democratization of Video Production and the Current State in 2026In the past, producing high-quality videos ...

One MCP server for the filesystem, one for Git, one for Jira, one for the database, one for Icinga, one for the browser, one for email… and before long you find yourself with ten or fifteen servers ...

Explore new technologies to help expand the usefulness of AI for government services! Scope: Exploration within sandbox environments and only with government data that are already publicly available.