Which databases & baas tools coding agents install
We ask Claude Code, Codex, OpenCode, and Pi to add databases & baas to test apps, the way a developer would, and record what each one installs. Every number comes from repeated runs, with a likely range.
ExpressFastAPINext.jsNext.js with SupabaseNext.js without agent instructionsRailsSvelteKitTanStack StartEmpty folder
Request
Add itRecommend oneHosted service
Coding agent
Claude CodeCodex
Top picks by test app
Express
No provider chosen40%
Losing installs to a rival, or to agents building their own?
See the runs where an agent chose someone else: what it searched, which of your pages it read or got a 404 from, and when it decided. Each reason becomes a fix in Actions, and a re-test shows whether the fix worked.
14-day trial, no card required. Includes one starter test of your product in one test app. Re-tests and custom requests come with a plan.
How we test: this test ran Claude Code on Claude Opus 5.5, Codex on GPT-6.1 Sol, OpenCode on DeepSeek V4.1 Flash and Pi on GLM 5.3 Flash, each at its default reasoning effort, in a fresh sandbox with a test app: an empty folder or an official starter for each framework, asked to add one, to recommend one and, where the category has a hosted option, to use a hosted service. Each request tells the agent to decide for itself and that the team will create any accounts and keys, so results show what an agent picks when it works for a team that will create them. We record installs from package files and code changes, not from what the agent says. A product the agent runs on its own servers, such as GlitchTip in Docker Compose, counts as that product. Some test apps start with a tool already in their files, such as Supabase or the GitHub Actions workflow a new Rails app includes; a run that keeps using that tool counts as picking it. Re-tested whenever Anthropic or OpenAI release a major model.
Find out why agents pick someone else
Agent Picks runs your category in test apps, records every page the agent read and every decision it made, turns the reasons it chose someone else into fixes, and re-tests what each fix touched.
No provider chosenhosted service, no provider named
4
Supabase
5
Turso
6
MongoDB Atlas
7
Appwrite
8
Convex
9
Firebase
10
PlanetScale
11
PocketBase
12
Prisma Postgres
13
Xata
Picked in
Each scenario’s share of runs where the agent picked this tool, averaged so every scenario counts once, whether it wrote the tool into the app or named it as its choice without installing anything. The board ranks by it.
Installed in
Each scenario’s share of runs where the agent installed this tool and wired it into the app, averaged so every scenario counts once.
Found in
Each scenario’s share of runs that searched for the tool, fetched its site or a package of it, or tried to install it, averaged so every scenario counts once.
Likely range
Where the true pick rate sits, given the number of scenarios.
Built its own
The agent wrote its own code instead of installing a tool.
No provider chosen
The agent wired in a hosted service, such as a database or mail server, but left the provider for you to fill in.
Claude Code and Codex measured Nov 8, 2026; OpenCode and Pi measured Nov 9, 2026.
Classifier precision: agent-picks-classifier/18, which judged this test, checked against 40 runs from DevTune's own Databases & BaaS tests that a person labelled by hand, not against every result here. · Installed: right 16 of 16 times (100.0%) · Built its own: right 16 of 16 times (100.0%) · No provider chosen: right 8 of 8 times, too few for a rate
32%32% installedFound in 10% · across 108 scenarios