gnomon
Manages a vault of articles and talk pages using an article/talk split pattern, enabling reading, editing, and librarian operations on the vault through MCP.
README
Gnomon
An article/talk split for the LLM wiki pattern.
A small structural addition to Karpathy's LLM Wiki pattern and a particular way of framing what the LLM is doing in it. The addition is the article/talk split that Wikipedia has had since the early 2000s: every article has a paired talk page. The framing is that the LLM works as a wiki gnome — substrate-tender, not author or compiler.
The load-bearing case is the one where multiple people and multiple LLM agents work on the same body of thinking and it has to stay legible across actors and time — who said what, what's settled, where the disagreement lives. Single-author note-taking is the easy case; gnomon is calibrated for the hard one. Said plainly: the example vault in this repo has so far had one human and one gnome in it. The multi-actor case is what the conventions are built for, and what the talk pages demonstrate in miniature — two kinds of author disagreeing, deciding, and leaving the trail — but it hasn't yet carried a full team, and that test matters.
For the full statement, read vault/articles/gnomon.md. The conventions are in vault/AGENTS.md. This README orients someone arriving cold; everything substantive lives in those two files.
The idea in one paragraph
An article holds the current best statement of something. Its paired talk page holds the working-out — disputes, drafts, decisions, signed and dated. The two are different kinds of writing and they get different treatment: articles are the settled corpus and stay small and trustworthy; talk is where volume accumulates. The LLM tends the structure (typos, links, frontmatter), surfaces issues on talk, and drafts on talk before any substantive edit — but it does not author articles, resolve disputes, or flatten the tensions the talk pages exist to hold. That restraint is the whole point: it keeps the settled corpus something a human was willing to sign, even as the agents generate faster than anyone can read.
What's in here
This repo is the tool plus a small example vault — the vault content here is a working demonstration of gnomon documenting itself.
vault/articles/— the example vault's articles. Concept articles and source articles, all in one flat namespace. Type is in frontmatter, not in the path.vault/talk/— paired talk pages, same basenames as the articles. Real working pages: disputes, drafts, decisions about what's settled.vault/sources/— raw external material referenced by source-type articles.vault/AGENTS.md— the schema. Read by Claude Code, Cursor, Aider, and any agent that follows theAGENTS.mdconvention. The LLM operating against the vault reads this file and follows its conventions. §4 is the gnome's operational checklist and is load-bearing.mcp/— MCP server (TypeScript). Vault operations grouped by what they touch: reading, originating, editing along a graduated model (janitorial → bold → authorial), posting to talk, librarian operations (compaction, sampling-as-audit, on-demand navigation), plus lint. One server can serve several vaults via repeated--vault name=path.site/— static site generator. Renders the vault to browseable HTML where each page shows the article and its talk page together — which is what makes the structural commitment visible.
Running
Requires Node 20+.
npm install
npm run lint # lint the example vault
npm run site # build the static site into vault/_site/
npm run site:serve # build and serve at localhost:8000
npm run mcp # start the MCP server (stdio) against the bundled vault
To use the MCP server with Claude Desktop or Claude Code, register it against a vault:
{
"mcpServers": {
"gnomon": {
"command": "npx",
"args": ["tsx", "/absolute/path/to/gnomon/mcp/server.ts", "--vault", "gnomon=/path/to/your/vault"]
}
}
}
The vault path can be this repo's vault/ (to operate against the example content) or any other directory laid out per AGENTS.md. Point --vault at as many vaults as you like.
Reading order
If you're new and have ten minutes:
vault/articles/gnomon.md— what gnomon isvault/articles/article-talk-split.md— the structural commitmentvault/articles/wiki-gnome.md— the LLM's role inside the splitvault/talk/gnomon.md— what a talk page actually looks like in use
If you have an hour:
vault/AGENTS.md— the full schema, in particular §4 (the gnome's operational checklist)vault/articles/karpathy-llm-wiki.md— what gnomon takes from and adds to Karpathy's patternmcp/server.ts— the MCP operations
Adapting
The whole thing is meant to be copied and adapted. Keep AGENTS.md and the articles/ + talk/ + sources/ layout; replace the example content with your own; point the MCP at the new vault. Different vaults can carry different schemas — edit AGENTS.md to match what your vault actually needs. Conventions, not ontology.
What this is and isn't
It isn't a startup, a product, a hosted service, or a position paper dressed as a tool. It's a studio piece in the idea-file-plus-reference-implementation mode, articulating one structural move and one framing move with an artifact small enough to read in an hour and adapt in a day.
The project enters Karpathy's conversation rather than opening a new one.
License
MIT. Take it, adapt it, run it. If you find something useful or notice something wrong, the talk pages are where that goes.
推荐服务器
Baidu Map
百度地图核心API现已全面兼容MCP协议,是国内首家兼容MCP协议的地图服务商。
Playwright MCP Server
一个模型上下文协议服务器,它使大型语言模型能够通过结构化的可访问性快照与网页进行交互,而无需视觉模型或屏幕截图。
Magic Component Platform (MCP)
一个由人工智能驱动的工具,可以从自然语言描述生成现代化的用户界面组件,并与流行的集成开发环境(IDE)集成,从而简化用户界面开发流程。
Audiense Insights MCP Server
通过模型上下文协议启用与 Audiense Insights 账户的交互,从而促进营销洞察和受众数据的提取和分析,包括人口统计信息、行为和影响者互动。
VeyraX
一个单一的 MCP 工具,连接你所有喜爱的工具:Gmail、日历以及其他 40 多个工具。
graphlit-mcp-server
模型上下文协议 (MCP) 服务器实现了 MCP 客户端与 Graphlit 服务之间的集成。 除了网络爬取之外,还可以将任何内容(从 Slack 到 Gmail 再到播客订阅源)导入到 Graphlit 项目中,然后从 MCP 客户端检索相关内容。
Kagi MCP Server
一个 MCP 服务器,集成了 Kagi 搜索功能和 Claude AI,使 Claude 能够在回答需要最新信息的问题时执行实时网络搜索。
e2b-mcp-server
使用 MCP 通过 e2b 运行代码。
Neon MCP Server
用于与 Neon 管理 API 和数据库交互的 MCP 服务器
Exa MCP Server
模型上下文协议(MCP)服务器允许像 Claude 这样的 AI 助手使用 Exa AI 搜索 API 进行网络搜索。这种设置允许 AI 模型以安全和受控的方式获取实时的网络信息。