Enter Her: A Debugging Layer for Claude Code
HuggingFace just dropped Her (हेर), an open-source tool that sits between you and Claude Code and records everything. Every prompt. Every response. Every token usage. The tool, available now on HuggingFace Spaces, offers a web UI that replays sessions like a video timeline. You can search, filter, and export logs. Think of it as a dashcam for your AI coding assistant. It captures the raw JSON stream and renders it in a clean interface — timestamps, token counts, even the model's internal reasoning traces if available. The initial release supports Claude Code v0.4 and above.
Why Claude Code Needed a Detective
Claude Code is powerful but opaque. Developers have been flying blind when things go wrong. You send a prompt, get a response, but what really happened? Was it a bad prompt, a context window issue, or a hallucination? Prior to Her, debugging meant manual logging or trusting Anthropic's minimal analytics. The community has been begging for visibility. Her is the first serious answer. Since last year, forums have filled with frustrated users trying to reproduce spurious code changes. Her gives them a forensic tool.
What Her Means for Developers
This is more than a nice tool. If you're building on Claude Code, Her fundamentally changes your feedback loop. You can now see exactly why your code changed. It encourages prompt engineering as a disciplined practice. Honestly, the most interesting part isn't the UI — it's the potential for cost optimization. Seeing 10,000 tokens wasted on irrelevant context? That's money burning. Her also makes it easier to comply with audits, since you have a full session log. One concrete example: a developer on Hacker News reported cutting his Claude Code bill by 40% after identifying long, redundant system prompts.
What's Still Missing in Her
Her is not perfect. It logs locally by default, raising questions about sensitive data in logs. It doesn't yet support multi-session comparisons or automated analysis — you still have to read the logs yourself. And while it claims minimal overhead, early users report 5-10% latency increase. That's non-trivial for interactive use. The big unknown: will Anthropic embrace this or try to block it? For now, it's a side-channel workaround using the CLI's unofficial output mode. No guarantees it won't break with a future update.