Software

Hugging Face Launches Her, a Tool for Analyzing Claude Code Sessions

Hugging Face introduces Her, a tool that analyzes Claude Code sessions, offering insights into tool usage and risky actions within 4,000-line JSON files.

Image: Hugging Face

Hugging Face has launched Her, a new tool designed to analyze Claude Code sessions. The tool processes the trace files generated during these sessions, offering insights into tool usage and risky actions. Users can upload a session file, and Her will investigate the data, reconstructing what happened in plain English. It flags risky moves, such as deployments and production changes, and traces them back to the exact turn where they occurred. The tool also identifies which tools, subagents, and MCP servers were used, providing a detailed breakdown of the session. The analysis is grounded in best practices from Anthropic and the community, offering suggestions without asserting findings. The tool is built to stay silent when there's nothing worth saying, ensuring that only relevant insights are provided. Source: huggingface

Her operates without calling any third-party AI APIs, using the Nemotron-Mini-4B-Instruct model running on the Space’s own GPU via ZeroGPU. The model is used solely to write English explanations and propose softer suggestions, without asserting findings. The evaluation engine is purely deterministic, ensuring that the analysis remains consistent and trustworthy. The session data is uploaded only to a private, auto-deleted namespace belonging to the user’s run, ensuring data privacy. The model’s output does not change with updates, maintaining the integrity of the analysis. The tool also includes a built-in copilot, allowing users to ask questions about tool usage and receive answers based on the session trace. Source: huggingface

The development of Her began as an operator’s view, featuring a journey graph where each query is a node sized by cost. The tool evolved after feedback from friends, leading to the addition of an executive report and a tool database. The frontend is a React app served off a Gradio server, with the deterministic engine handling the forensics and Nemotron managing the prose. The tool is designed to assist users when Claude Code sessions become complex, offering a reliable way to analyze and understand the session data. Source: huggingface

Key points

Her reads .jsonl files from Claude Code sessions to reconstruct what happened in plain English.
The tool flags risky moves such as deploys, config and production changes, and secrets.
Her identifies CLI tools used during a session and provides a one-line blurb for each.
The model Nemotron-Mini-4B-Instruct runs on the Space’s own GPU via ZeroGPU.
Session data is uploaded only to a private, auto-deleted namespace belonging to the user’s run.
The evaluation engine is purely deterministic, ensuring consistent and trustworthy analysis.
Her does not assert findings, only suggesting improvements when a named, fixable pattern is detected.

Source: Hugging Face Read the original →

WRITTEN BY

Theo Almeida

AI Software & Developer Tools

Theo covers AI software, developer tools, frameworks, and the platforms builders use every day.