mcp-local-rag MCP — Сценарии использования, Установка & Живое демо

Зачем использовать

Ключевые функции

Fully local — embeddings computed with on-device model, zero network after first run
Hybrid search: semantic (embeddings) boosted with keyword matching
Supports PDF, DOCX, TXT, Markdown, HTML — common personal-knowledge formats
Single command setup via npx mcp-local-rag; also usable as a CLI
CPU-only — works on any modern laptop

Живое демо

Как выглядит на практике

local-rag.replay ▶ готово

0/0

Установка

Выберите клиент

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

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

Откройте Claude Desktop → Settings → Developer → Edit Config. Перезапустите после сохранения.

~/.cursor/mcp.json · .cursor/mcp.json

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

Cursor использует ту же схему mcpServers, что и Claude Desktop. Конфиг проекта приоритетнее глобального.

VS Code → Cline → MCP Servers → Edit

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

Щёлкните значок MCP Servers на боковой панели Cline, затем "Edit Configuration".

~/.codeium/windsurf/mcp_config.json

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

Тот же формат, что и Claude Desktop. Перезапустите Windsurf для применения.

~/.continue/config.json

{
  "mcpServers": [
    {
      "name": "local-rag",
      "command": "npx",
      "args": [
        "-y",
        "mcp-local-rag"
      ]
    }
  ]
}

Continue использует массив объектов серверов, а не map.

~/.config/zed/settings.json

{
  "context_servers": {
    "local-rag": {
      "command": {
        "path": "npx",
        "args": [
          "-y",
          "mcp-local-rag"
        ]
      }
    }
  }
}

Добавьте в context_servers. Zed перезагружается автоматически.

claude mcp add local-rag -- npx -y mcp-local-rag

Однострочная команда. Проверить: claude mcp list. Удалить: claude mcp remove.

Сценарии использования

Реальные сценарии: mcp-local-rag

Build a private RAG over your downloaded research papers and PDFs

👤 Researchers, students, knowledge workers ⏱ ~30 min beginner

Когда использовать: You've hoarded hundreds of PDFs in ~/Documents/papers and want to actually use them — 'what did that paper say about attention decay?'

Предварительные требования

PDFs or docs on disk — Any folder of files — recursive ingest supported

Поток

Ingest the folder

Ingest everything under ~/Documents/papers into local-rag. Skip files larger than 50MB.✓ Скопировано

→ Per-file ingest log + 'indexed N files' summary
Ask questions

Across my papers, what do they say about positional encoding in long-context transformers? Cite the source file and page if possible.✓ Скопировано

→ Synthesized answer with source file citations
Refine search

Just give me the top 5 passages most relevant to 'ring attention', raw — don't summarize.✓ Скопировано

→ Ranked passage list

Итог: Every paper you've ever downloaded is now queryable by topic — permanent upgrade to your reading life.

Подводные камни

Scanned PDFs have no extractable text — Run an OCR pass first (ocrmypdf) before ingesting
First index of 1000+ files is slow (CPU embeddings) — Leave it running overnight; incremental re-ingest is fast

Сочетать с: filesystem

Query confidential contracts / HR docs without leaking to any cloud

👤 Legal ops, HR, compliance ⏱ ~20 min intermediate

Когда использовать: Documents are too sensitive for OpenAI/Claude cloud embeddings. You need search but can't send content anywhere.

Поток

Ingest

Ingest /secure/contracts/*.pdf into local-rag.✓ Скопировано

→ Files indexed locally; confirm no network call was made
Query

Which contracts have an auto-renewal clause longer than 12 months?✓ Скопировано

→ List of candidate contracts with the clause quoted

Итог: Searchable private corpus with nothing leaving the machine.

Подводные камни

Claude answers still go to Anthropic — the embeddings are local but conversation isn't — If answers must also be local, run with a local LLM via Ollama or LM Studio instead of cloud Claude

Сочетать с: filesystem

Index your codebase docs and Markdown for instant 'where is this explained?' lookups

👤 Engineers on large repos ⏱ ~15 min beginner

Когда использовать: Your team wiki / design docs / ADRs are scattered across Markdown files and grep doesn't find things by meaning.

Поток

Ingest the docs tree

Ingest all .md files under ./docs. Re-ingest when I ask.✓ Скопировано

→ Indexed file list
Ask architecture questions

Where's the rationale for using event sourcing instead of CRUD? I remember an ADR about it.✓ Скопировано

→ Pointer to the specific ADR file + quoted rationale

Итог: Team knowledge you can search by concept, not filename.

Подводные камни

Docs drift — old ADRs contradict current state — Re-ingest before important searches; delete_file for deprecated docs

Сочетать с: filesystem

Комбинации

Сочетайте с другими MCP — эффект x10

local-rag + filesystem

Watch a folder, re-ingest files when they change

Every time a file under ~/Notes changes, re-ingest it into local-rag.✓ Скопировано

local-rag + firecrawl

Scrape a docs site then feed to local-rag for offline querying

Crawl docs.example.com, save each page as Markdown, then ingest all of them into local-rag.✓ Скопировано

local-rag + playwright

Capture JS-rendered pages and ingest their extracted text

Open this SPA, grab the rendered HTML, ingest_data it into local-rag with the URL as source.✓ Скопировано

Инструменты

Что предоставляет этот MCP

Инструмент	Входные данные	Когда вызывать	Стоимость
ingest_file	path: str \| path[]	Add one or more files to the index	CPU only
ingest_data	html: str, source_url?: str	Add a raw HTML blob — useful after scraping	CPU only
query_documents	query: str, top_k?: int	Main retrieval call — use before answering user questions	free
list_files		See what's indexed	free
delete_file	path: str	Remove a stale/irrelevant file from the index	free
status		Sanity check index size	free

Стоимость и лимиты

Во что обходится

Квота API: None — all local
Токенов на вызов: Query results 500-3000 tokens depending on top_k
Деньги: Free. One-time ~90MB model download.
Совет: Set top_k to 5-8 for most questions; going higher wastes tokens without improving answers.

Безопасность

Права, секреты, радиус поражения

Хранение учётных данных: None — no API keys

Исходящий трафик: Zero after model download. Your docs never leave the machine.

Single-user — no multi-user isolation. Don't share the index directory.
Changing embedding model requires wiping the DB and re-ingesting everything.

Устранение неполадок

Частые ошибки и исправления

First query is slow / seems to hang

Embedding model is downloading on first run (~90MB). Subsequent calls are fast.

Проверить: Check ~/.cache/mcp-local-rag for the model file

PDF ingest returns 0 chunks

PDF is likely scanned (image-only). Run ocrmypdf input.pdf output.pdf first.

Проверить: pdftotext input.pdf -

Results feel irrelevant

Pure semantic search struggles with short queries. Add more keywords. The hybrid search boosts them already.

Out of memory on large PDFs

Split the PDF first, or raise Node heap: NODE_OPTIONS=--max-old-space-size=8192

Альтернативы

mcp-local-rag в сравнении

Альтернатива	Когда использовать	Компромисс
Chroma MCP / Qdrant MCP	You want a real vector DB with multi-user, scaling, metadata filters	More setup, usually requires a running server
OpenAI Assistants file_search	You're OK sending documents to OpenAI's cloud	Not local, costs per token, but zero setup and more accurate
ChatGPT Projects / Claude Projects file upload	Small document set (<20 files) and you use the hosted chat	Not an MCP; can't be scripted

Ещё

Ресурсы

📖 Читать официальный README на GitHub

🐙 Открытые задачи

🔍 Все 400+ MCP-серверов и Skills