Hugging Face fingerprints which coding agent you run

Your coding agent wears a name badge, and it hands it over on every Hugging Face request. The huggingface_hub Python library detects which agent harness is driving it — Claude Code, Codex, Cursor, Copilot — and stamps that name into the User-Agent header on every call to the Hub. It is documented, it is deliberate, and the aggregate numbers ship as a public dataset. Most teams running AI coding agents have no idea it happens.

What actually happened

This is not a leak or a bug. It is a feature Hugging Face documents openly: "When huggingface_hub detects it is running inside a registered harness, it reports it via the user agent on Hub requests."

The mechanics are worth reading. _http_user_agent() in _headers.py calls detect_agent() and, when it returns a name, appends ; agent/<name> to the header. Detection itself reads environment variables: the universal AI_AGENT and AGENT, plus per-tool variables pulled from a registry the library fetches from /api/agent-harnesses and caches for 24 hours.

That registry is a public file — agent-harnesses.ts — with 26 entries today. Claude Code is spotted via CLAUDECODE or CLAUDE_CODE. Codex via CODEX_SANDBOX, CODEX_CI, CODEX_THREAD_ID. Cursor CLI via CURSOR_AGENT. GitHub Copilot via COPILOT_MODEL and siblings. Replit via REPL_ID. Zed via ZED_TERM.

The August 2026 dataset, updated September 7, puts Claude Code at 46.53% of agent-attributed requests and 38.71% of distinct authenticated users. Codex takes 17.52% and 23.34%. Cursor CLI takes 14.04% and 5.38%.

Here is the detail nobody mentions: HF_HUB_DISABLE_TELEMETRY=1 does not suppress the agent token. In _headers.py that flag gates whether your torch version gets appended. detect_agent() runs outside the branch.

Why this matters for your business

Two things, and they point opposite directions.

If you pull models, know what you disclose. The header carries your tool choice, and when a token is set it rides alongside an authenticated user. Not your code, not your repo — but "this account runs Claude Code" is metadata you did not decide to publish. If you build under NDA for clients, that belongs in the same bucket as any other outbound telemetry: inventoried, or turned off. There is no supported off switch, so the control is environmental — strip the detection variables from the shell that runs the job, or make the call over plain HTTP, which the library never tags.

If you cite the numbers, cite them honestly. Attribution only covers traffic through huggingface_hub and the hf CLI. Raw HTTP is invisible. Unregistered tools bucket into unknown. This is a floor on agent traffic, not a market-share census, and anyone quoting it as one is selling something.

Sources: Hugging Face Hub docs — Agents, huggingface/agent-usage dataset.