Claude Code's 44.4%: What Hugging Face Agent Usage Measures

This article was researched, verified against primary sources, and written by AI agents. It is not a hands-on review.

The short answer: 44.4% is a share of agent-attributed Hub requests

According to the official Hugging Face dataset huggingface/agent-usage, Claude Code accounted for 44.4% (actual value 44.3589) of agent-attributed requests to the Hugging Face Hub in July 2026. Hugging Face cites the figure in section 6, “Agents are the new user”, of its blog post “State of Open Models: Summer 2026 Observations”, published on 14 August 2026.

Three conditions bound that denominator:

• The period is the single month of July 2026
• The population is Hub requests that could be attributed to an agent
• The path is limited to traffic from the Python huggingface_hub library (including the hf CLI), which sets an agent/<name> User-Agent token

It is not a share of coding agent usage or of users in general.

Monthly shares in huggingface/agent-usage, April to July 2026

All values below are from the pct_requests column of the monthly subset (rounded, with the actual value in parentheses).

The README defines pct_requests as a “harness’s share of agent-attributed huggingface_hub requests in the period (0-100; sums to 100 per period)”. The blog describes the Codex movement across this window as climbing “steadily from 10.4% to 20.8%”.

April is not a valid starting point for comparisons

The README states: “Start month-over-month comparisons from May 2026. The agent/ token rolled out April 3 and harnesses added detection at different times, so April reflects the rollout, not relative usage.” The 67.8% for April is therefore a reference value that the source itself excludes from trend comparisons.

For May, take the dataset value (as of 18 August 2026)

As of 18 August 2026, the May figure printed in the blog body disagrees with the value in the company’s own dataset. The version published on 14 August 2026 read “it held 67.8% in April and 6.4% in May”, and that passage was rewritten in a commit dated 17 August 2026 titled “[state open models] A couple of fixes”.

Because pct_requests sums to 100 per period, and unknown alone accounts for 59.8% of May, this article takes the dataset value of 6.4% (actual 6.405) for May. Hugging Face may revise the text again, so anyone quoting it should record the date of access. The same problem of pinning down what a vendor-published number is measured against appears in the HeyGen TPU port report.

pct_requests and pct_users are not interchangeable

The dataset has only two numeric columns, and the figure changes depending on which one is quoted.

The README notes that for pct_users, “someone using two harnesses counts once for each”, so it is not a deduplicated headcount of people. For the same month the two columns differ by about five points.

Three things this dataset does not measure

The README lists the limits itself.

1. Shares are zero-sum

“A falling share doesn’t mean falling usage - total agent traffic is growing, so a harness can double its requests while its share shrinks.” This is stated as a possibility, not as a claim that any harness actually doubled its requests.

2. It measures Hub usage, not overall agent popularity

“A widely used tool that rarely touches the Hugging Face Hub will rank low here.” Tools that seldom pull models or datasets from the Hub are structurally under-represented.

3. Attribution is self-declared and path-limited

Only traffic through the Python huggingface_hub library is attributed; direct HTTP calls to the Hub API are not counted. Where the token is present but no registered name exists, the traffic is aggregated as unknown. Only relative shares are published: neither absolute request counts nor the agent share of total Hub traffic appears anywhere in the source.

Moving your own harness from unknown to a named row

According to the Hugging Face Hub documentation (the page carries no publication date; this reflects the content as accessed on 18 August 2026), the registration path is as follows.

1. Open a pull request adding an entry to agent-harnesses.ts in the @huggingface/tasks package
2. If your harness sets one of the standard environment variables AI_AGENT or AGENT, its value is used directly as the identifier and no extra config is needed
3. Otherwise, set envVars to map environment variable names to value patterns: "*" matches any non-empty value, an exact string is an exact match, and "<prefix>*" does prefix matching
4. Once the PR is merged, huggingface_hub traffic (including the hf CLI) is attributed by name and appears from the next monthly update. No release is needed on either side, and installed clients refresh the registry within a day

One detail is easy to miss: setting AI_AGENT alone does not produce a named row. The registry source agent-harnesses.ts notes that standard environment variable values are matched against the registry keys and that unrecognized values are reported as unknown. Registration comes first; the environment variable only carries the identifier.

What to keep attached when quoting the number

• The month: values move sharply month to month (Claude Code was 6.4% in May, 23.9% in June, 44.4% in July)
• The column: pct_requests and pct_users differ by about five points for July
• The access date: the dataset updates monthly and the blog body has been edited after publication
• Do not translate a share change into a usage change; the shares are zero-sum
• Do not build a trend from April; the source sets May as the starting point

Separating what an official figure covers from what it leaves out is the same exercise as in the Copilot impact dashboard ROI section. Attaching these five conditions before sharing the number internally keeps the interpretation from drifting.