datasette-llms-txt

This plugin generates llms.txt documentation for a Datasette instance. Large language models and people can then find out which data is available and how to query it.

The plugin reads the schema of your databases when a request arrives. You do not build or update any files. It serves Markdown that lists every database, table, column, and primary key. It also gives the JSON and SQL endpoints that query the data. If you write descriptions in Datasette metadata, the plugin uses them without change.

Installation

Install this plugin in the same environment as Datasette:

datasette install datasette-llms-txt

Usage

After you install the plugin, it adds three endpoints:

For example, with a demo.db file:

datasette demo.db
curl http://127.0.0.1:8001/llms.txt
curl http://127.0.0.1:8001/demo/llms.txt
curl http://127.0.0.1:8001/demo/people/llms.txt

The plugin also adds a link to /llms.txt in Datasette's top-right menu.

Downloading data

Each per-table llms.txt also shows how to download the data, not only how to query it:

• CSV export: /<db>/<table>.csv
• Full table, streamed: /<db>/<table>.csv?_stream=on — every row, no row limit
• Full table as JSON: request /<db>/<table>.json?_size=max, then follow the next_url field in each response
• Whole database file: /<db>.db — the plugin shows this only when the database is immutable and the allow_download setting is on

The plugin shows the streamed-CSV and database-file options only when the Datasette settings allow_csv_stream and allow_download permit them. Because of this, the generated docs never link to an endpoint that returns a 403 error.

Metadata

The plugin prefers descriptions from your Datasette metadata over text that it builds from the schema. It shows the instance title and description, the per-database and per-table description, the per-column descriptions, and the source and license.

A table's description has two uses. It becomes the summary line for that table in the index and the per-database listings. It also becomes the blockquote at the top of the table's own llms.txt. If you do not give a description, the plugin generates a summary like "N rows, columns: …". For example:

title: My Data
description: Public records, refreshed nightly.
source: Example Org
source_url: https://example.com
databases:
  demo:
    tables:
      people:
        description: One row per registered person.
        columns:
          email: Primary contact address.

Configuration

Set the options under the plugin's key in your metadata or config file:

plugins:
  datasette-llms-txt:
    sample_rows: 5

• sample_rows (default 5) — the number of example rows to show in each per-table llms.txt. Set it to 0 to leave out all sample data. This helps when the rows are sensitive.

Development

cd datasette-llms-txt
uv venv
uv pip install -e '.[test]'
uv run pytest

Or use the Makefile: make install then make test.

Publishing

Build a release and upload it to PyPI with uv:

uv build                       # produces dist/*.whl and dist/*.tar.gz
uv publish --token <pypi-token>   # or set UV_PUBLISH_TOKEN

You can also run make build and then make pypi. The make pypi target reads UV_PUBLISH_TOKEN from the environment. It is better to publish with the GitHub Actions publish.yml workflow than to upload by hand. That workflow runs when you publish a GitHub Release, and it uses trusted publishing.