panasenco/mcp-sqlite

Provide useful data to AI agents without giving them access to external systems. Compatible with Datasette for human users!
Features
- AI agents can get the structure of all tables and columns in the SQLite database in one command - sqlite_get_catalog.
- The catalog can be enriched with descriptions for the tables and columns using a simple YAML or JSON metadata file.
- The same metadata file can contain canned queries to the AI to use.
Each canned query will be turned into a separate MCP tool sqlite_execute_main_{tool name}.
- AI agents can execute arbitrary SQL queries with sqlite_execute.
Quickstart using Visual Studio Code
1. Install uv.
2. Install Visual Studio Code if you don't already have it.
Turn on GitHub Copilot.
3. Open this repo in VS Code.
Open a GitHub Copilot agent mode chat.
Check the available tools - you should see MCP Server: sqlite_sample with three available tools.
4. You should be able to ask Copilot in agent mode a question like "Get Titanic survivors of age 28" and get a response.
5. Use the sample MCP configuration file mcp.json and the sample metadata file
titanic.yml as a starting point for your own configuration.
Interactive exploration with MCP Inspector and Datasette
The same database and metadata files can be used to explore the data interactively with MCP Inspector and Datasette.
| MCP Inspector | Datasette |
| ------------- | --------- |
| | |
| | |
MCP Inspector
Use the MCP Inspector dashboard to interact with the SQLite database the same way that an AI agent would:
1. Install npm.
2. Run:
Datasette
Since mcp-sqlite metadata is compatible with the Datasette metadata file, you can also explore your data with Datasette:
Compatibility with Datasette allows both AI agents and humans to easily explore the same local data!
MCP Tools provided by mcp-sqlite
- **sqlite_get_catalog()**: Tool the agent can call to get the complete catalog of the databases, tables, and columns in the data, combined with metadata from the metadata file.
In an earlier iteration of mcp-sqlite, this was a resource instead of a tool, but resources are not as widely supported, so it got turned into a tool.
If you have a usecase for the catalog as a resource, open an issue and we'll bring it back!
- **sqlite_execute(sql)**: Tool the agent can call to execute arbitrary SQL. The table results are returned as HTML.
For more information about why HTML is the best format for LLMs to process, see Siu et al.
- **{canned query name}({canned query args})**: A tool is created for each canned query in the metadata, allowing the agent to run predefined queries without writing any SQL.
Usage
Command-line options
Metadata
Hidden tables
Hiding a table with hidden: true will hide it from the catalog returned by the MCP tool sqlite_get_catalog().