Which services win most often in the Armature data?

Stripe takes 9 out of 10 payment selections, Neon leads databases with a 66% win rate, Amazon S3 holds 45% of file-storage picks, and Resend and Postmark lead email at 35.6% and 27.4%. Armature notes the repository's programming language shifts these outcomes noticeably.

Do the agents search the web before choosing a tool?

They differ sharply. Armature found Codex uses web search in 94% of sessions, often with targeted search operators, and Cursor searches in roughly two-thirds of sessions. Claude Code searches in only about 30% of runs, relying mostly on knowledge it already carries, which is one reason the three agents disagree so often.

Why does a well-known service like PayPal lose?

Armature's data separates being mentioned from being selected. PayPal appeared 139 times across the measured sessions and was never the tool an agent actually wrote into the code. Brand recall inside a model's training data does not translate into the final integration decision, which is where the revenue is.

How often do agents skip third-party tools entirely?

Claude Code builds an in-house solution roughly twice as often as its competitors — 19% of sessions against 10% for Codex and Cursor. For a vendor, that means a share of every prompt is lost not to a rival service but to the agent deciding the feature is simple enough to write from scratch.

Armature · 2026-09-03 · major

Coding agents agree on tools only 42% of the time — 16,893 sessions

Armature ran 16,893 sessions across Claude Code, Codex and Cursor on 75 repositories, then recorded which third-party service each agent reached for. The three agents picked the same tool in only 42% of cases.

Armature chart of which third-party tools Claude Code, Codex and Cursor select — Armature

Armature measured 16,893 coding-agent sessions to see which third-party services Claude Code, Codex and Cursor actually install.

Quick facts

Published	September 3, 2026
Publisher	Armature
Sessions	16,893 measured; 5,292 published in wave 1
Repositories	75 across 10 programming languages
Prompt variations	1,163
Agents tested	Claude Code, Codex, Cursor
Agreement rate	42% on tool choice

What is it?

Published September 3, 2026, the Armature study asks a question nobody had measured at scale: when you tell a coding agent to add payments, a database or email to a project, which vendor does it choose? The team ran 16,893 sessions across Claude Code, Codex and Cursor, over 75 repositories in 10 programming languages, using 1,163 prompt variations. The first wave publishes 5,292 valid sessions.

How does it work?

Each run gave an agent a repository and a feature request, then recorded the service it wrote into the code. Claude Code, Codex and Cursor land on the same tool in only 42% of runs, and they get there differently: Codex searches the web in 94% of sessions, Cursor in roughly two-thirds, and Claude Code in about 30%, leaning instead on what it already knows. Claude Code also writes its own implementation instead of pulling a dependency in 19% of runs, against 10% for the others.

Why does it matter?

For anyone shipping a developer tool, agents are becoming a distribution channel with its own rules. Armature's numbers show clear winners — Stripe takes 9 of 10 payment picks, Neon leads databases at a 66% win rate, Amazon S3 holds 45% of file storage, and Resend and Postmark lead email at 35.6% and 27.4%. Being well known is not the same as being chosen: PayPal was mentioned 139 times across sessions and selected zero times.

Who is it for?

developer-tool founders, DevRel teams, and engineers building agent workflows

Frequently asked questions

Which services win most often in the Armature data?: Stripe takes 9 out of 10 payment selections, Neon leads databases with a 66% win rate, Amazon S3 holds 45% of file-storage picks, and Resend and Postmark lead email at 35.6% and 27.4%. Armature notes the repository's programming language shifts these outcomes noticeably.
Do the agents search the web before choosing a tool?: They differ sharply. Armature found Codex uses web search in 94% of sessions, often with targeted search operators, and Cursor searches in roughly two-thirds of sessions. Claude Code searches in only about 30% of runs, relying mostly on knowledge it already carries, which is one reason the three agents disagree so often.
Why does a well-known service like PayPal lose?: Armature's data separates being mentioned from being selected. PayPal appeared 139 times across the measured sessions and was never the tool an agent actually wrote into the code. Brand recall inside a model's training data does not translate into the final integration decision, which is where the revenue is.
How often do agents skip third-party tools entirely?: Claude Code builds an in-house solution roughly twice as often as its competitors — 19% of sessions against 10% for Codex and Cursor. For a vendor, that means a share of every prompt is lost not to a rival service but to the agent deciding the feature is simple enough to write from scratch.

Try it

Read the first-wave results at https://armature.tech/blog/which-tools-coding-agents-install