Do Claude Code, Codex and Cursor pick the same tools?

If you measure how coding agents choose your product and you only measure one agent, you know about half of your market.

We ran the same requests through Claude Code, Codex and Cursor, inside the same repositories, 5,292 times. Here is how often they agreed.

The headline

The full matrix

What agreement looks like

The eight categories where all three agree share a property: one product is very far ahead.

When a product holds this much of a category, every agent reaches for it without much deliberation. It is in the training data, in the templates, and in the repositories.

Agreement is not a sign that the agents are thinking alike. It is a sign that there is nothing left to decide.

What disagreement looks like

The categories where they differ are the ones with no runaway leader. In search, the top product holds 17%. In agent frameworks, 14%. In AI gateways, 21%.

In those categories the agent is genuinely choosing, and small differences in how each agent works produce different answers.

If you are deciding where to spend as a result of this, which coding agent should you optimise for? works through it.

Why they disagree

Three causes, in order of size.

They search the web at very different rates

This is the biggest single difference between the two, and it is enormous. Codex is reading the current web in half its sessions. Claude Code is almost always working from what it already knows plus what is in the repository.

The practical consequence: a page you publish reaches Codex often and Claude Code almost never. If your users run Claude Code, content is a weak lever and the repository is a strong one.

They differ on building it themselves

Claude Code reaches for a hand-written implementation nearly twice as often as the other two. If you sell in a category where hand-written code is a plausible answer, Claude Code is a harder market, and the argument you have to win is not "why us over them" but "why a product at all".

Their training data and defaults differ

Every model carries a different set of habits. In authentication, Claude Code put WorkOS AuthKit first while both others put Auth0 first. In email, Claude Code put Postmark first while both others put Resend first. Those are stable patterns across hundreds of runs each, not noise.

The self-preference question

In agent frameworks, Cursor put the Cursor SDK first, while Claude Code and Codex both put the Vercel AI SDK first.

It is worth being careful here. The Cursor SDK is a real product in that category and a reasonable choice inside the Cursor environment. We also wrote the agent framework prompts to name the makers in alternating order, with a linter that checks it, precisely because we expected this question to come up.

So the honest statement is: in one category, one agent put its own maker's product first while the other two did not. That is one observation. It is worth watching rather than a proven general rule, and we will keep measuring it.

What this means if you sell software

Measure per agent, always. A blended install share is an average across markets that behave differently. It will tell you that you hold 22% of a category while hiding that you hold 40% with one agent and 5% with another. The fix for those two situations is not the same.

Match the lever to the agent. Content and search visibility move Codex. The repository, the templates and the training data move Claude Code. Find out which agent your users run before choosing where to spend.

Treat a split category as an opportunity. If the three agents disagree, the category is not settled. Nothing is locked in and the ordering can be moved. If all three agree and the leader holds 60%, you are not going to change that with a blog post this quarter, and you should be looking for the requests where the leader fits badly.

Watch the in-house rate per agent. Losing to Claude Code writing its own implementation is a completely different problem from losing to a competitor, and it needs different content: an argument that the problem is harder than it looks, with specifics.

How this was measured

5,292 judged sessions across 18 categories. Real command line tools at pinned versions, in sandboxes, inside 51 realistic codebases with real lock files and replayed git histories. Requests written as four kinds of person, in several wordings. A simulated project owner approves or questions each recommendation before any code is written. A judge from a model family that builds none of the three agents reads every session blind.

Three copies of the judge agreed with each other 97.9% of the time.

Every run is published and can be replayed on the sector boards.