Lab · Study · 600 answers · 2 Nov 2026
Which tools do AIcoding agents install?
Ask an AI to add payments to your app and it picks the product for you. We asked Claude, GPT and Gemini 600 times and counted what they installed.
From the lab
30/30
Sentry and Playwright: installed in every answer.
The question
Who picks
the tools now?
More and more code is written by AI agents. When the agent types the install command, it chooses the product, and most people keep what it chose. For a developer tool, being picked by AI is the new search ranking.
- 600answers, each one a real build task
- 20categories, from auth to vector search
- 3models: Claude, GPT and Gemini
- 0products named in any prompt
What we found · 20 categories
The tool AI
installs most.
Installs out of 30 answers in each category. A long bar means the agents agree; a short one means the race is open.
-
Error monitoring
Sentry
30/30
-
E2E testing
Playwright
30/30
-
Product analytics
PostHog
28/30
-
Payments
Stripe
24/30
-
Postgres hosting
Neon
20/30
-
UI components
shadcn/ui
19/30
-
Object storage
AWS S3
17/30
-
ORM
Drizzle
16/30
-
Transactional email
Resend
16/30
-
Hosting
Vercel
12/30
-
Background jobs
BullMQ
12/30
-
Cron and workflows
A cron library
12/30
-
Logging
Grafana stack
12/30
-
Search
Meilisearch
11/30
-
Headless CMS
Sanity
11/30
-
Feature flags
PostHog flags
10/30
-
Vector database
pgvector
10/30
-
LLM observability
Langfuse
10/30
-
Realtime
Socket.IO
10/30
-
Auth
Auth.js
8/30
What surprised us
Named is not
installed.
- 4 → 0Talked about, never pickedSupabase came up 4 times for Postgres hosting and was installed 0 times. Radix UI: named 11 times, installed 0.
- 40 / 20 / 0The models disagreeFor auth, Claude chose Auth.js in 40% of answers, GPT in 20%, and Gemini not once: it went with Clerk.
- 16 vs 5An old favourite losesDrizzle beat Prisma 16 to 5 as the database layer. Claude never installed Prisma.
- 30 / 30Some races are overSentry for errors and Playwright for tests were installed in every single answer.
What it means for you
If AI doesn't
pick you, buyers won't see you.
The same shift is coming for every product, not only developer tools. People ask ChatGPT what to buy, and it answers with two or three names. We check whether yours is one of them, in Arabic and English, and fix it.
Method
How we
ran it.
What did you ask?
Five build tasks per category, written the way a developer would ask, such as "add login to my Next.js app". No prompt names a product. Each answer had to end with the exact install command.
How many answers per category?
Thirty: five prompts, two samples each, on Claude, GPT and Gemini. Twenty categories make 600 answers. The run was on 2 November 2026.
Is this what an agent in my editor does?
Close, not the same. These runs go through the models' APIs, so they are repeatable. An agent inside your project also sees your files, and may stay with what you already use.
Can I use the numbers?
Yes. The data is open (CC BY 4.0) on installrate.com, with every prompt and answer.
From the lab
89%
of the words, heard right by the best engine. The weakest got 32%.
Start here
Tell the lab
where you are.
An idea, a launch, search or AI: one call tells you the next step.
We reply within one working day.

