Which tools do Claude Code, Codex, and Cursor choose?
Armature
Armature ran 16,893 coding-agent tool picks across 75 repos. Claude Code, Codex and Cursor often disagree — and context changes the winner.
Based on reporting by Armature — read the original for the full story.
Summary, retelling and take written by AI under human oversight; images are AI-generated illustrations. How we work · Report an error
Armature says it ran the biggest test yet of how coding agents choose third-party tools: 16,893 runs, 75 repositories, 1,163 prompt variations and three agents — Claude Code, Codex and Cursor — actually implementing the choices. The setup wasn’t a toy benchmark. It used fake company names, fake git histories, fake API keys and real lockfiles, with sandboxes rotated across E2B, Blaxel and Daytona.
The point was simple: if agents are starting to act like the people who pick the services inside a codebase, what do they actually trust? Armature built scenarios around four kinds of users, from vibe-coders to enterprise engineers, and even added a simulated human step so the agent could recommend a tool first, then get approval to install it. That mattered. Without that back-and-forth, the agents leaned harder toward building things themselves or defaulting to platform-native choices.
The first thing Armature found is that these systems do not think alike. Cursor based its choices on web results in about two-thirds of sessions. Codex searched the web in 94% of cases, and often used operators like site: to narrow in on trusted pages. Claude Code leaned much more on its prior knowledge and only searched the web about 30% of the time, though in newer areas like sandboxes it searched far more often. Across all three, the same tool won in only 42% of cases.
Repository context also mattered a lot. Give the same request to four repos in four languages and you can get four different winners. Resend led on TypeScript, Sendgrid on Python, Postmark on Go and Azure ACS on Java. Vercel won on TypeScript repos and in every Next.js case, but never on Python, where Render took over.
And then there’s the gap between being mentioned and actually being chosen. Paypal showed up 139 times in payment-provider sessions and never won. Adyen was mentioned 175 times and picked three times. LangChain was cited 194 times and chosen four times. Supabase was the most mentioned database at 242 references, yet Neon still came out ahead. The lesson here is pretty blunt: if vendors want agent traffic, they need to be the easiest credible answer, not just the loudest name on the page.
My take — AI-written commentary, not fact-checked reporting
Agents are becoming the new middlemen, which is great news for whoever writes the cleanest docs and worst for anyone still praying for brand loyalty. This is the part the hype crowd keeps missing: the model doesn’t care about your logo, it cares about whichever page makes the choice feel safe. Very rude, very efficient, very internet.
Read more about this at: Armature