node-sqlite-mcp

An MCP server that lets an AI agent query SQLite database files on the local machine.

It uses Node's built-in node:sqlite module, so there is no native module to compile and no system SQLite to install — the only thing you need on the host is a recent Node.

Requirements

• Node.js 22.16 or newer (which ships node:sqlite unflagged). Node 24+ is recommended.

Check with node --version.

Related MCP server: SQLite MCP Server

Install

The server is published to npm and is installed into your agent with add-mcp:

npx add-mcp node-sqlite-mcp

add-mcp detects your installed agents (Claude Code, Claude Desktop, Cursor, VS Code, …) and registers the server as a local stdio command. Nothing is cloned or installed globally — the command is fetched and run on demand.

You can also run it directly to sanity-check it:

npx -y node-sqlite-mcp

It speaks MCP over stdio and waits for a client; there is nothing interactive to see.

Tools

Every tool takes the database as an argument — path should be an absolute path to a .db / .sqlite file. The database is always opened read-only.

Output format

Results come back as TSV: a tab-separated header row of column names, one line per row, then a blank line and a row count. This is used instead of JSON because it does not repeat column names on every row, keeping the payload small.

• NULL is rendered as an empty field.

• Tabs, carriage returns, newlines and backslashes inside values are escaped (\t, \r, \n, \\). Other control characters are escaped as \xNN (e.g. NUL as \x00).

• BLOBs are rendered as a SQLite hex literal, e.g. x'deadbeef'.

• The describe_table not_null and primary_key columns are booleans (true/false).

Read-only

The SQLite connection is opened read-only, so INSERT / UPDATE / DELETE and schema changes fail at the engine level — not merely by convention. The query tool runs a single statement per call and does not truncate results, so a broad SELECT * on a large table can return a very large payload; the tool description tells the agent to add a LIMIT and select only needed columns.

Security note

This is a local, trust-the-user tool. It will open any SQLite file the operating-system user running the server can already read — there is no allowlist or directory confinement.

Development

npm install      # installs deps and builds via the "prepare" script
npm run build    # compile TypeScript to dist/

Source is TypeScript in src/; the published package ships compiled JavaScript in dist/ (the bin entry point). Consumers never run a build step.

Releases & publishing

Releases are automated with Release Please, driven by Conventional Commits:

• Every push to main opens or updates a standing release PR that bumps the version in package.json / package-lock.json and updates CHANGELOG.md, derived from the commits since the last release (fix: → patch, feat: → minor, a ! or BREAKING CHANGE: → major).

• Merging that release PR tags the version, creates a GitHub Release, and triggers npm publish (on-push-main.yml → workflow-call.release.yml).

So the normal flow is: land Conventional-Commit PRs on main, then merge the release PR when you want to cut a version — no manual tag or npm publish. Publishing requires an NPM_TOKEN repository secret: an npm automation token with publish rights to the node-sqlite-mcp package.

To publish by hand instead (unscoped packages publish publicly by default):

npm publish

Commit messages

This repo uses Conventional Commits. Because PRs are merged (not squashed), every commit lands on main verbatim and is what Release Please parses — so each commit in a PR is linted against the convention (on-pr.yml → workflow-call.lint-pr.yml, config in .commitlintrc.json). Common types: feat:, fix:, docs:, chore:, refactor:, test:, ci:. A feat!: prefix or a BREAKING CHANGE: footer marks a breaking change.

License

MIT