Key Takeaways
- Trace Commons lets Claude Code users donate anonymized session traces to an open CC-BY-4.0 dataset on Hugging Face.
- Run
/donate-traceafter open-source work to share how you solved problems — without exposing secrets or paths.
What Changed — A Public Dataset for Agent Traces
Trace Commons (https://trace-commons-web.hf.space/) is a new initiative to build an open, CC-BY-4.0-licensed dataset of coding-agent session traces on Hugging Face. The goal is simple: democratize the data that currently only companies like Anthropic and OpenAI can see.
"Every coding-agent session is a record of how these tools really work — and today that data goes to a few companies," the project states. "Donate yours instead, anonymized on your machine, to an open dataset anyone can study and build on."
The dataset is hosted on Hugging Face and is downloadable by anyone. It already supports 50+ coding agents including Claude Code, Codex, pi, and opencode.
What It Means For You — Your Traces, Anonymized and Public
Your Claude Code session traces contain the reasoning path — not just the diff. "A diff shows where you landed; a trace shows how you got there," the project explains. "That reasoning is what's locked inside a few companies today."
For Claude Code users, this means:
- Your workflow becomes research data. Researchers, open-source tool builders, and competitors can study how real developers solve problems with Claude Code.
- No privacy risk. Paths, usernames, and secrets are stripped on your machine before anything leaves. You review the redacted trace before confirming.
- Attribution optional. If you're logged into Hugging Face, you can attribute your trace. Otherwise, it's anonymous.
- Maintainers review first. The trace opens as a pull request that a project maintainer reviews before anything goes public.
This is a direct response to the closed nature of agent training data. Claude Code itself runs on models like Claude Opus 4.6 and Claude Sonnet 4.6 — but the traces of how those models actually behave in real coding sessions are locked inside Anthropic. Trace Commons opens that loop.
Try It Now — One Command to Start Donating
npx skills add trace-commons-ai/donate-trace
This installs the skill into your Claude Code session. After any open-source session, run:
/donate-trace
The skill:
- Confirms the repo is public (no proprietary code leaks)
- Anonymizes paths, usernames, and secrets locally
- Shows you exactly what was removed
- Opens a pull request for maintainer review before anything goes public
For other agents like pi, the command is /skill:donate-trace.
When to Use It
- After solving a tricky bug — your trace shows the reasoning chain others can learn from
- After completing an open-source PR — the maintainer can review the trace alongside your code
- After a refactor — traces reveal how Claude Code reasons about existing codebases
What Gets Shared
Each donation becomes one public row in the dataset. You can browse existing contributions directly on Hugging Face: https://huggingface.co/datasets/trace-commons/traces
Why This Matters for Claude Code Users
Coding agents improve on data their makers can see. Right now, only Anthropic sees how Claude Code behaves in the wild. An open dataset means:
- Researchers can study failure modes and build better benchmarks
- Tool builders can train alternative models on real agent behavior
- The community can audit how these tools actually work
"Open infrastructure is built, not given — one session at a time," the project concludes. Your /donate-trace command is a brick in that foundation.
Related: If you're interested in cross-agent memory sharing, check out our coverage of MACCHA, the file-based cross-agent brain that makes Claude Code remember.
Source: trace-commons-web.hf.space
