SQLite MCP — Сценарии использования, Установка & Живое демо

Зачем использовать

Ключевые функции

One-file database — no server to run
Read and write SQL — useful for both analysis and ETL
Schema introspection (tables, columns, indexes)
Stored as a single file you can cp, git, or attach to email
Excellent fit for personal/CLI/analytics workflows

Живое демо

Как выглядит на практике

sqlite.replay ▶ готово

0/0

Установка

Выберите клиент

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

{
  "mcpServers": {
    "sqlite": {
      "command": "uvx",
      "args": [
        "mcp-server-sqlite",
        "--db-path",
        "/data/sample.db"
      ]
    }
  }
}

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

~/.cursor/mcp.json · .cursor/mcp.json

{
  "mcpServers": {
    "sqlite": {
      "command": "uvx",
      "args": [
        "mcp-server-sqlite",
        "--db-path",
        "/data/sample.db"
      ]
    }
  }
}

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

VS Code → Cline → MCP Servers → Edit

{
  "mcpServers": {
    "sqlite": {
      "command": "uvx",
      "args": [
        "mcp-server-sqlite",
        "--db-path",
        "/data/sample.db"
      ]
    }
  }
}

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

~/.codeium/windsurf/mcp_config.json

{
  "mcpServers": {
    "sqlite": {
      "command": "uvx",
      "args": [
        "mcp-server-sqlite",
        "--db-path",
        "/data/sample.db"
      ]
    }
  }
}

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

~/.continue/config.json

{
  "mcpServers": [
    {
      "name": "sqlite",
      "command": "uvx",
      "args": [
        "mcp-server-sqlite",
        "--db-path",
        "/data/sample.db"
      ]
    }
  ]
}

Continue использует массив объектов серверов, а не map.

~/.config/zed/settings.json

{
  "context_servers": {
    "sqlite": {
      "command": {
        "path": "uvx",
        "args": [
          "mcp-server-sqlite",
          "--db-path",
          "/data/sample.db"
        ]
      }
    }
  }
}

Добавьте в context_servers. Zed перезагружается автоматически.

claude mcp add sqlite -- uvx mcp-server-sqlite --db-path /data/sample.db

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

Сценарии использования

Реальные сценарии: SQLite

Analyze a CSV/JSON dump by loading it into SQLite

👤 Analysts, engineers exploring exported data ⏱ ~15 min beginner

Когда использовать: Someone sent you a CSV with 200k rows and the question 'which segment converts best?' — too big for spreadsheets, too small for a real DB.

Предварительные требования

Source file on disk — Save as .csv or .json under a working folder
An empty SQLite file path — Choose a location like /tmp/analysis.db; the MCP will create it

Поток

Create the table and load

Create a table signups in /tmp/analysis.db matching the columns of /data/signups.csv. Load all rows. Tell me the row count.✓ Скопировано

→ Table created, row count matches file
Explore the schema

What columns exist? For each, what's the value distribution (top 5 distinct values for categoricals; min/max/avg for numerics)?✓ Скопировано

→ Per-column profile
Answer the actual question

Group by signup_source. For each, compute: total signups, conversion rate (signups with completed_onboarding=true / total). Sort by conversion rate.✓ Скопировано

→ Decision-grade table with the SQL shown

Итог: Defensible answers in 5 minutes, with a .db file you can re-query as new questions come up.

Подводные камни

CSV columns auto-typed wrong (numbers as TEXT) — After load, run PRAGMA table_info(signups) and CAST or recreate columns with explicit types if needed
Date strings don't sort/compare correctly as TEXT — Store dates as ISO 8601 (YYYY-MM-DDTHH:MM:SSZ) so lexicographic = chronological; or use julianday() for math

Сочетать с: filesystem · antv-chart

Inspect and edit a personal app's SQLite database

👤 Devs building CLI tools, journal apps, or local-first software ⏱ ~10 min beginner

Когда использовать: You're building a local-first app, you want to see what's in the DB without writing a CLI for it.

Поток

Survey the schema

List every table in /Users/me/Library/Application Support/MyApp/data.db. For each, show schema and row count.✓ Скопировано

→ Inventory of the live app DB
Investigate a row

Find the user record where email = '[email protected]'. Show the row and any related rows in other tables (orders, sessions).✓ Скопировано

→ Full picture of one user's data
Fix bad data

There's a stuck order in 'pending' state for that user from 2 days ago. Update it to 'cancelled'. Show the SQL before running.✓ Скопировано

→ SQL preview before mutation, then row updated

Итог: App debugging without writing throwaway SQL scripts.

Подводные камни

App may have the DB locked open with WAL mode — If you get 'database is locked', stop the app, or query the WAL-merged read-only snapshot via ?mode=ro&immutable=1

Сочетать с: filesystem

Build deterministic test fixtures from a sample of prod data

👤 Engineers writing integration tests ⏱ ~25 min intermediate

Когда использовать: You want repeatable test data that resembles prod but is small and safe.

Поток

Sample anonymized rows

From /prod-export/orders.db, sample 100 rows from orders covering each status. Anonymize names and emails.✓ Скопировано

→ Sample with anonymized PII
Save as a fixture file

Write the sampled rows to /test/fixtures/orders.db as a fresh SQLite file. Include schema.✓ Скопировано

→ New fixture file created
Verify against your test loader

Run my test suite (npm test) — does it pick up the new fixture? If not, what's the first failing test?✓ Скопировано

→ Tests run; failures pinpointed

Итог: Realistic fixtures that don't drift from real data shapes.

Подводные камни

Anonymization that breaks referential integrity — Anonymize join keys consistently (same hash) across tables; never randomize per-row

Сочетать с: filesystem · github

Analyze a SQLite-backed log/event file

👤 Engineers debugging CLI tools or apps that log to SQLite ⏱ ~10 min beginner

Когда использовать: Many modern tools (homebrew, some browsers, app caches) store state in SQLite. You want to query them.

Поток

Confirm it's the right file

Open ~/Library/Application Support/SomeApp/cache.db. List tables and a sample of recent rows.✓ Скопировано

→ Recognizable schema confirms you have the right file
Find the answer

How many entries does the cache hold per source domain? Top 20.✓ Скопировано

→ Aggregation result
Optionally clean up

Delete entries from domains that haven't been accessed in 90 days. Show count first, ask before deleting.✓ Скопировано

→ Preview, confirmation, then delete

Итог: Answers about app behavior with no need for a built-in 'stats' command.

Подводные камни

Modifying an app's live DB while the app is running can corrupt it — Always close the app first, or work on a copy of the .db file

Сочетать с: filesystem

Комбинации

Сочетайте с другими MCP — эффект x10

sqlite + filesystem

Read a CSV from disk and load into SQLite for analysis

Read /data/orders.csv with filesystem MCP, infer types, and load into /tmp/analysis.db as table orders via sqlite MCP.✓ Скопировано

sqlite + antv-chart

Query a SQLite DB and chart the results

From /tmp/analysis.db, get monthly signups for 2026. Render as a bar chart via antv-chart.✓ Скопировано

sqlite + github

Analyze data, write findings into a GitHub Issue

Run my churn analysis on /tmp/users.db. Create a GitHub Issue in acme/analytics summarizing the top 3 findings with the SQL appendix.✓ Скопировано

Инструменты

Что предоставляет этот MCP

Инструмент	Входные данные	Когда вызывать	Стоимость
list_tables	none	First step in any session	free
describe_table	table_name: str	Inspect a single table's schema	free
read_query	query: str (SELECT only)	Run a SELECT — safe by default	free
write_query	query: str (INSERT/UPDATE/DELETE)	Mutate data — gated; requires explicit consent in most clients	free
create_table	query: str (CREATE TABLE ...)	DDL — create or alter schema	free
append_insight	insight: str	Add a finding to the session memo (used by some clients to build a report)	free

Стоимость и лимиты

Во что обходится

Квота API: Unlimited — local
Токенов на вызов: Schema queries: small. Result sets scale with row count — always LIMIT for exploratory queries
Деньги: Free
Совет: Add LIMIT 100 to every exploratory query and only remove it when you know what you're getting back.

Безопасность

Права, секреты, радиус поражения

Хранение учётных данных: No credentials. The DB file is whatever path you launch with via --db-path.

Исходящий трафик: None from the server. Query results ship to your LLM provider as context.

Никогда не давайте: never point at a file holding sensitive data unless you intend the model to see it

The MCP can run write_query — destructive. Some clients ask before each mutation; some don't. Check your client's policy.
If a SQLite file holds PII, treat the entire file as confidential.

Устранение неполадок

Частые ошибки и исправления

database is locked

Another process (often the app that owns the DB) holds the lock. Close that process or copy the .db file and query the copy.

Проверить: lsof <db file>

no such table: X

Wrong DB file or schema not what you think. Run list_tables to see what's actually there. Check the launch arg --db-path in your MCP client config.

datatype mismatch / unexpected NULL

SQLite is dynamically typed — a column declared INTEGER can hold TEXT. Use CAST(col AS INTEGER) defensively, or fix at load.

Disk image is malformed

DB corrupted, often from killing a process during a write. Try sqlite3 file.db .recover > out.sql and rebuild from the dump.

Альтернативы

SQLite в сравнении

Альтернатива	Когда использовать	Компромисс
Postgres MCP	Multi-user concurrent access, networked DB, or you're already on Postgres	Needs a server; Postgres MCP is read-only by design
DuckDB (via shell)	Same one-file model but for OLAP-shaped analytics with much faster scans	No first-party MCP yet; columnar so different perf characteristics
dbHub	You need one MCP for SQLite + Postgres + MySQL + others	Newer; less battle-tested

Ещё

Ресурсы

📖 Читать официальный README на GitHub

🐙 Открытые задачи

🔍 Все 400+ MCP-серверов и Skills