Hugging Face dataset: Claude Code tops Hub agent requests

A quiet Hugging Face release this week is worth a moment of your attention if you sell, build, or buy coding agents. The company published a public dataset that tracks which coding-agent harnesses are actually calling the Hugging Face Hub, and the July 2026 numbers put Claude Code on top with 44.4% of agent-attributed requests and 39.5% of distinct users. Codex sits second at 20.8% of requests and 21.2% of users, with Cursor's CLI at 5.8% and 5.9%.

The reason to care is that most claims about which coding agent is 'winning' have so far come from vendor decks and screenshots. This is one of the few cross-vendor snapshots that comes from real request logs rather than surveys, published as a versioned Parquet dataset with a monthly subset covering April through July 2026 and a daily one running to more than two thousand rows.

The methodology is narrower than the top-line makes it sound, and Hugging Face is unusually blunt about that. Attribution comes from an `agent/<name>` User-Agent token that harnesses set through the `huggingface_hub` Python library, so a tool that talks to the Hub over raw HTTP is invisible, and a tool whose maintainers have not yet registered themselves in the harness list shows up bucketed as `unknown`. Only relative shares are published; raw counts are withheld. Because the shares sum to 100 per period, a falling percentage is not the same as falling usage, and the company explicitly flags April 2026 as an unreliable rollout month.

The honest caveat is that Hub-specific traffic is not the same as overall coding-agent popularity, and what the file does not give you is the absolute request volume or how much of general Hub traffic is now agent-driven in the first place. Take the ranking as directional, not settled.

Still, the direction is the interesting part. If you build a coding-agent product, there is now a public leaderboard you can appear on by registering. If you buy one, you finally have a share signal that is not a vendor slide.