mcp-documentation-server MCP — 使用场景, 安装 & 实时演示

为什么要用

核心特性

Local-first: no cloud DB, no external calls unless you enable Gemini
Parent-child chunking balances precision and context
Hybrid full-text + vector search
Built-in web dashboard on port 3080 for drag-and-drop uploads
Supports .txt, .md, and .pdf out of the box

实时演示

实际使用效果

documentation-server.replay ▶ 就绪

0/0

安装

选择你的客户端

~/Library/Application Support/Claude/claude_desktop_config.json · Windows: %APPDATA%\Claude\claude_desktop_config.json

{
  "mcpServers": {
    "documentation-server": {
      "command": "npx",
      "args": [
        "-y",
        "mcp-documentation-server"
      ],
      "_inferred": true
    }
  }
}

打开 Claude Desktop → Settings → Developer → Edit Config。保存后重启应用。

~/.cursor/mcp.json · .cursor/mcp.json

{
  "mcpServers": {
    "documentation-server": {
      "command": "npx",
      "args": [
        "-y",
        "mcp-documentation-server"
      ],
      "_inferred": true
    }
  }
}

Cursor 使用与 Claude Desktop 相同的 mcpServers 格式。项目级配置优先于全局。

VS Code → Cline → MCP Servers → Edit

{
  "mcpServers": {
    "documentation-server": {
      "command": "npx",
      "args": [
        "-y",
        "mcp-documentation-server"
      ],
      "_inferred": true
    }
  }
}

点击 Cline 侧栏中的 MCP Servers 图标，然后选 "Edit Configuration"。

~/.codeium/windsurf/mcp_config.json

{
  "mcpServers": {
    "documentation-server": {
      "command": "npx",
      "args": [
        "-y",
        "mcp-documentation-server"
      ],
      "_inferred": true
    }
  }
}

格式与 Claude Desktop 相同。重启 Windsurf 生效。

~/.continue/config.json

{
  "mcpServers": [
    {
      "name": "documentation-server",
      "command": "npx",
      "args": [
        "-y",
        "mcp-documentation-server"
      ]
    }
  ]
}

Continue 使用服务器对象数组，而非映射。

~/.config/zed/settings.json

{
  "context_servers": {
    "documentation-server": {
      "command": {
        "path": "npx",
        "args": [
          "-y",
          "mcp-documentation-server"
        ]
      }
    }
  }
}

加入 context_servers。Zed 保存后热重载。

claude mcp add documentation-server -- npx -y mcp-documentation-server

一行命令搞定。用 claude mcp list 验证，claude mcp remove 卸载。

使用场景

实战用法： mcp-documentation-server

Make a new framework's docs queryable by your AI

👤 Devs adopting a new library ⏱ ~20 min beginner

何时使用： Official docs are huge; you want Claude to answer with grounded citations.

前置条件

mcp-documentation-server installed — npx -y @andrea9293/mcp-documentation-server

步骤

Ingest the docs

Upload the library's docs .md files to the dashboard at http://localhost:3080.✓ 已复制

→ Files processed into chunks
Ask targeted questions

search_documents for 'how to configure middleware' — give me the top 3 chunks with source paths.✓ 已复制

→ Cited excerpts
Ask grounded synthesis

Given those chunks, write the minimum viable config for middleware in this framework.✓ 已复制

→ Working config backed by cited doc lines

结果： A personal docs assistant that cites its sources.

注意事项

PDFs with scanned images aren't OCR'd — Pre-OCR with tools like ocrmypdf before upload
Huge doc sets without Gemini give noisy embeddings — Optional GEMINI_API_KEY unlocks higher-quality semantic search

搭配使用： filesystem

Turn an internal wiki export into a RAG source

👤 Teams with markdown-export-friendly wikis ⏱ ~25 min beginner

何时使用： You've exported your Notion/Confluence content as Markdown and want AI access.

步骤

Bulk-ingest via process_uploads

process_uploads on ./wiki-export/ — process every .md.✓ 已复制

→ Document count per folder
Full-scope search

search_all_documents: 'deployment runbook' — top 5.✓ 已复制

→ Ranked list

结果： A local, private, searchable wiki.

Build a personal research-paper library

👤 Researchers, students ⏱ ~30 min beginner

何时使用： You download papers and want them queryable instead of piled in Downloads/

步骤

Drop PDFs in

Upload all PDFs in ~/Papers/ to the documentation server.✓ 已复制

→ Papers chunked and indexed
Ask across corpus

search_documents: 'attention variants with lower quadratic cost' — return authors + years.✓ 已复制

→ Cited excerpts

结果： A local mini-perplexity over your own paper collection.

组合

与其他 MCP 搭配，撬动十倍杠杆

documentation-server + filesystem

Automate ingest from a watched folder

Every time a new PDF lands in ~/Papers/Inbox, process_uploads it into the documentation server.✓ 已复制

documentation-server + swarmvault

Compare: documentation-server is quick-ingest; swarmvault builds a structured wiki

Ingest my research PDFs into both systems; compare retrieval quality on the same query.✓ 已复制

工具

此 MCP 暴露的能力

工具	输入参数	何时调用	成本
add_document	title, content, metadata?	Programmatic ingest	free (local embeddings)
list_documents	(none)	See what's indexed	free
get_document	id	Retrieve a specific doc	free
delete_document	id	Pruning	free
search_documents	query, top_k?	Query within a specific doc set	free
search_all_documents	query, top_k?	Global RAG query	free
get_context_window	chunk_id	Expand a narrow hit into broader context	free
search_documents_with_ai	query	One-shot answer synthesis	Gemini call (needs key)
process_uploads	path?: str	Batch import from the uploads folder	free

成本与限制

运行它的成本

API 配额: None if local; Gemini usage if GEMINI_API_KEY set
每次调用 Token 数: Search returns 500-3000 tokens depending on top_k
费用: Free; Gemini is paid per-call if enabled
提示: Skip Gemini for exploratory work — local embeddings are good enough for known-item lookups.

安全

权限、密钥、影响范围

凭据存储： GEMINI_API_KEY (optional) in env

数据出站： Local only unless Gemini is enabled; dashboard on port 3080

Dashboard on :3080 has no auth — bind to localhost only.
Document content sits in the MCP_BASE_DIR as JSON; treat it like any local data.

故障排查

常见错误与修复

Port 3080 in use

Set WEB_PORT env var to another port.

验证： lsof -i :3080

PDF parse error

Password-protected or scanned PDFs fail. Remove password or run OCR first.

验证： Try a plain PDF

search returns empty

Check documents ingested: list_documents. If empty, re-run process_uploads.

验证： list_documents

替代方案

mcp-documentation-server 对比其他方案

替代方案	何时用它替代	权衡
swarmvault	You want a structured wiki + knowledge graph, not just search	Heavier; more upfront setup
Cloud RAG (Pinecone, Weaviate)	You need team-shared, scale	Paid; data leaves your machine
llm-context.py	You want per-task context, not persistent doc retrieval	Different problem

mcp-documentation-server