Self-hosted Postgres
16of 36 runs installed it
- share
- 44%
- named only
- 0
- rank in category
- #2
Postgres vs SQLite: what coding agents actually pick
SQLite 20, Postgres 16, in 36 database runs. Opus 5.5 went Postgres instead.
36 database runs in Claude Code, the prompt never names a tool. updated Oct 18, 2026 · all database results
Postgres vs SQLite, side by side
SQLite file
picked more20of 36 runs installed it
- share
- 56%
- named only
- 0
- rank in category
- #1
where it splits
runs that installed each one, out of runs scored. a highlighted row went the other way from the total.
by modelSelf-hosted PostgresSQLite file
- Opus 5.5flips8of 124of 12
- Sonnet 5.53of 129of 12
- Haiku 5.55of 127of 12
by starterSelf-hosted PostgresSQLite file
- Next.js7of 1811of 18
- Python9of 189of 18
every database result
36 runs, all models. most common: SQLite file, 20 of 36.
- SQLite file20 of 3656%
- Self-hosted Postgres16 of 3644%
get told if Self-hosted Postgres catches up
new models get the same database prompts. one email with what moved. free. read issue 1
Postgres vs SQLite, questions
Postgres vs SQLite, which one do coding agents pick?
SQLite file got installed in 20 of 36 database runs across Opus 5.5, Sonnet 5.5 and Haiku 5.5 (56%). Self-hosted Postgres got installed in 16.
does the claude model change the pick?
Opus 5.5: Postgres 8, SQLite 4, of 12. Sonnet 5.5: Postgres 3, SQLite 9, of 12. Haiku 5.5: Postgres 5, SQLite 7, of 12.
does the starter stack change the pick?
Next.js + TypeScript: Postgres 7, SQLite 11, of 18. Python + FastAPI: Postgres 9, SQLite 9, of 18.
how often did it name one without installing it?
named in the final message but not in the code: Postgres 0, SQLite 0.
how this was measured
Postgres is a database server. SQLite is a file next to the app. we asked Claude Code to add a real database to a blank app without naming one and counted what it wired in.
every run starts from the same blank starter, Next.js + TypeScript or Python + FastAPI. a pick is whatever ends up installed or wired into the code, read from the diff. the 3 prompts:
- “This app needs to store data that survives a restart. Set up a database for users and posts, connect the app to it, and show me how to run it locally.”
- “Add persistence to this project. I want a users table and a posts table, a connection configured through environment variables, and one endpoint that reads from the database.”
- “Right now everything is in memory and I lose it on reload. Move the data into a real database and wire up the connection.”
the database chart
other head to heads
- Resend vs SendGrid11 to 0
- Drizzle ORM vs Prisma12 to 0
- Sentry vs PostHog21 to 4
- Fly.io vs Render8 to 0
- Heroku vs Fly.io7 to 8
- Resend vs SMTP11 to 24
claude model comparison
- claude model comparison, all models
- Opus 5.5 vs Sonnet 5.5same lead in 7 of 10
- Opus 5.5 vs Haiku 5.5same lead in 7 of 10
- Sonnet 5.5 vs Haiku 5.5same lead in 10 of 10
how a pick gets counted: every run starts from the same blank starter repo, the prompt never names a tool, and a pick is whatever the agent installs or wires in, read from the diff. full rules on the methodology page, every run for database on the category page, raw data in runs.json.