给 AI Agent 的入口:一键安装我们的 Skill or MCP(free)
这个博客除了给人看,也对外提供 MCP(Model Context Protocol)服务——你的 AI agent(比如 Claude Code)可以直接连接过来,搜索、读取、订阅这里的文章,不用再靠爬网页解析 HTML。
方式一:直接注册成 MCP server(推荐)
如果你用的是 Claude Code,一行命令就能把这个博客接进去:
claude mcp add --transport http mushroom-blog https://blog.mushroom.cv/api/mcp
接上之后,search_posts/get_post/list_recent 会变成 Claude
能直接调用的原生工具,问"mushroom 博客最近发了什么"、"搜一下 mushroom 博客里关于 xxx 的文章",Claude
会自己去调用,不需要你再手动 curl。协议版本是 2025-11-25(不是最新的
2026-07-28,具体原因见 agent-feed/PLAN.md)。
方式二:安装 mushroom skill(更自然的触发方式)
如果你想要更简单的触发词——打 mushroom 就汇报最近文章,mushroom search 关键词
就搜索——装这个 skill:
mkdir -p ~/.claude/skills/mushroom
curl -s https://blog.mushroom.cv/agents/mushroom-skill.md -o ~/.claude/skills/mushroom/SKILL.md
装完重开一个 Claude Code 会话就能用。这个 skill 底层直接 curl 调 /api/mcp,不需要额外配置,
也不依赖方式一的 MCP 注册——两者互不冲突,都装也没问题。
方式三:纯文本,给任何能读网页的 agent
不想接协议、只想让 agent 读一份干净的文本索引:
- /llms.txt — 站点简介 + 按分类分组的文章列表
- /llms-full.txt — 同上,但每篇带完整正文
- /agent-feed-index.json — 结构化 JSON,程序直接解析用
这些服务是怎么来的
完整的设计决策(为什么选 MCP 不选 A2A/Nostr、协议版本为什么是 2025-11-25、两轮安全 review 改了什么) 记在仓库的 agent-feed/ 目录里, 公开可读。这套东西本身也在往"任何服务都能这样对 agent 开放"这个方向抽象,设计文档在 agent-feed/framework/。
Entry Point for AI Agents: One-Click Install Our Skill or MCP (Free)
Beyond being a website for humans, this blog also exposes an MCP (Model Context Protocol) server — your AI agent (e.g. Claude Code) can connect directly to search, read, and subscribe to articles here, instead of scraping and parsing HTML.
Option 1: Register as an MCP server (recommended)
If you use Claude Code, one command connects it:
claude mcp add --transport http mushroom-blog https://blog.mushroom.cv/api/mcp
Once connected, search_posts/get_post/list_recent become native
tools Claude can call directly — ask "what has the mushroom blog posted recently" or "search the mushroom
blog for xxx" and Claude will call them itself, no manual curl needed. Protocol version is
2025-11-25 (not the newest 2026-07-28 — see
agent-feed/PLAN.md for why).
Option 2: Install the mushroom skill (a friendlier trigger word)
For a simpler trigger — type mushroom for a recent-posts digest, or
mushroom search <query> to search — install this skill:
mkdir -p ~/.claude/skills/mushroom
curl -s https://blog.mushroom.cv/agents/mushroom-skill.md -o ~/.claude/skills/mushroom/SKILL.md
Works after restarting your Claude Code session. It calls /api/mcp via curl under the hood,
needs no extra config, and doesn't require Option 1's MCP registration — both can be installed together.
Option 3: Plain text, for any web-reading agent
Don't want to speak a protocol at all — just want a clean text index:
- /llms.txt — site summary + articles grouped by category
- /llms-full.txt — same, with full article bodies
- /agent-feed-index.json — structured JSON for programmatic use
How this came to be
The full design record (why MCP over A2A/Nostr, why protocol version 2025-11-25, what two rounds of adversarial security review changed) lives in the repo's agent-feed/ directory, publicly readable. This is also being generalized toward "any service can open up to agents this way" — design doc in agent-feed/framework/.