Stop Parsing JSON: Observe Claude Code Sessions with Her

Agent Observability: Her for Claude Code

Hugging Face Spaces now hosts Her, a deterministic forensic tool for Claude Code sessions. It addresses the “write-only” nature of .jsonl session traces by reconstructing agent behavior into plain English without requiring external API calls.

Technical Mechanics

• Local Inference: The tool runs entirely within the Space using Nemotron-Mini-4B-Instruct via ZeroGPU. No data leaves the private, auto-deleted namespace.
• Deterministic Core: The evaluation engine is purely deterministic. The LLM is strictly limited to generating prose and soft suggestions; it does not drive findings or alter metrics.
• Offline Tool Identification: A bundled database of Homebrew, npm, and PyPI packages allows offline identification of CLI tools, providing one-line blurbs for context.

Developer Utility

Her parses session files to flag high-risk actions such as production deploys, configuration changes, or secret exposure. It traces these events back to specific turns, showing token consumption by tool, subagent, or MCP server. Users can drop single files for session views or multiple files for project-wide queries. The built-in “Ask Her” copilot cites specific turns and opens the relevant tool call, enabling rapid root-cause analysis for failed or expensive runs.

Implementation Note

The frontend is a React app served via Gradio. For developers managing large-scale Claude Code deployments, this offers a self-hostable pattern for auditing agent behavior without leaking session data to third-party analytics services. It shifts the focus from raw JSON parsing to semantic understanding of agent decision paths.

Sources

• Huggingface: https://huggingface.co/blog/build-small-hackathon/her-blog