
基于 Git 支持的维基引擎搭配 MCP 服务器——构建知识可积累
概览
一个无头的代理维基引擎。24个MCP工具。一个Rust二进制文件。内部没有大型语言模型。 一个基于git支持的Markdown维基——可搜索、输入、图链接。可通过命令行访问,或通过任何兼容MCP的代理,或通过ACP的任何IDE访问。 大多数AI知识工具会检索并生成每个查询。每个答案都是一次性的——既不学,也不保留。问同一个问题两次,LLM会从零推理。 LLM-wiki 实现了另一种模式——由 Andrej Karpathy 引入的 (DKR): 在导入时处理源,而非查询时。LLM将每个源集成到维基中——更新概念页、创建源代码摘要、标记矛盾——并提交结果。知识随着每次添加而积累。 引擎纯粹是基础设施。它管理文件、git、全文搜索和图表结构。LLM始终是外部的——它通过MCP调用引擎工具,读取页面、写入页面并提交知识。
概述(AI 翻译)
一个无头的代理维基引擎。24个MCP工具。一个Rust二进制文件。内部没有大型语言模型。
积累积累的知识——而不是消散的答案。
一个基于git支持的Markdown维基——可搜索、输入、图表链接。访问地址 命令行,无论是从任何兼容MCP的代理,还是通过ACP从任何集成开发环境(IDE)操作。
RAG的问题
大多数AI知识工具会检索并生成每条查询。每个答案都是 一次性——学不到东西,也没被保留。同一个问题再问两次 以及从零开始推理LLM的逻辑。
LLM-wiki实现了另一种模式——动态知识库 (DKR),由安德烈·卡帕西介绍:
在导入时处理源,而非查询时。LLM 集成了两个 维基中的源代码——更新概念页,创建源代码摘要, 指出矛盾——并导致结果。知识与 每一个新增内容。
| 传统RAG | LLM-wiki(DKR) | |
|---|---|---|
| 知识构建时 | 查询时 | 摄入时 |
| 交叉引用 | 临时或遗漏 | 预构建的打字图 |
| 知识积累 | 重置每次查询 | 复合词随时间变化 |
本页由 AI 翻译,技术细节请以原文为准。
README
以下为原文 README 节选,完整内容请查看原文。
llm-wiki
A headless wiki engine for agents. 24 MCP tools. One Rust binary. No LLM inside.
Build knowledge that compounds — not answers that evaporate.
A git-backed Markdown wiki — searchable, typed, graph-linked. Accessible from the command line, from any MCP-compatible agent, or from any IDE via ACP.
The problem with RAG
Most AI knowledge tools retrieve and generate on every query. Each answer is disposable — nothing is learned, nothing is kept. Ask the same question twice and the LLM reasons from scratch.
llm-wiki implements a different pattern — the Dynamic Knowledge Repository (DKR), introduced by Andrej Karpathy:
Process sources at ingest time, not query time. The LLM integrates each source into the wiki — updating concept pages, creating source summaries, flagging contradictions — and commits the result. Knowledge compounds with every addition.
| Traditional RAG | llm-wiki (DKR) | |
|---|---|---|
| When knowledge is built | At query time | At ingest time |
| Cross-references | Ad hoc or missed | Pre-built, typed graph |
| Knowledge accumulation | Resets each query | Compounds over time |
| Audit trail | None | Git history per page |
| Data ownership | Provider systems | Your files, your git repo |
How it works
The engine is pure infrastructure. It manages files, git, full-text search, and graph structure. The LLM is always external — it calls the engine’s tools via MCP, reads pages, writes pages, and commits knowledge. Intelligence flows through skills, not the binary.
LLM agent
│
├── wiki_list(format: "llms") → all pages grouped by type
├── wiki_search("mixture of experts") → ranked results + facets
├── wiki_content_read("concepts/moe") → full page + backlinks
├── wiki_graph(root: "concepts/moe") → typed graph in Mermaid/DOT
├── wiki_suggest("concepts/moe") → pages worth linking
├── wiki_content_new("concepts/new-page") → scaffold + returns local path
├── [write directly to path] → no MCP round-trip
└── wiki_ingest(path: "concepts/") → validate, index, commit
A wiki page is a plain Markdown file with typed frontmatter:
---
type: concept
title: Mixture of Experts
status: active
confidence: 0.9
tags: [routing, scaling, efficiency]
sources:
- sources/switch-transformer-2021
- sources/mixtral-2024
concepts:
- concepts/sparse-routing
- concepts/scaling-laws
---
Sparse routing of tokens to expert subnetworks...
The engine validates frontmatter against a JSON Schema, extracts typed graph
edges from sources and concepts, and indexes everything in tantivy. The
graph is live the moment a page is committed.
Install
安装
This server does not publish a one-line install command.
Open the repository installation guide 新手?先看 mcpskill.org 教程 →配置
{
"mcpServers": {
"llm-wiki": {
"command": "llm-wiki",
"args": ["serve"]
}
}
}