Hugging Face Launches Her: The Forensics Tool for Claude Code
Bridging the gap between autonomous code generation and human observability.
Hugging Face has launched "Her," an open-source forensic tool designed to decode and analyze Claude Code sessions, providing developers with unprecedented visibility into autonomous agent behavior. By reconstructing complex JSONL traces into plain English, Her allows engineering teams to audit production changes, identify token waste, and verify critical security protocols in real-time. This release marks a pivotal moment for the developer community, transforming the "black box" of agentic workflows into a transparent, navigable narrative.
Key Details
Her serves as a specialized "detective" for AI-driven software development, addressing the "write-only" nature of complex agent traces that often span thousands of lines of raw data. In practice, most developers find it nearly impossible to manually audit a session that has spiraled into 4,000 lines of JSON code. Her solves this by providing a suite of analytical features:
- Automated Narrative Reconstruction: It converts raw .jsonl session files into a chronological narrative, making it easy to see exactly when and why an agent made a specific decision.
- Risk Identification and Flagging: The tool automatically flags high-risk activities, such as production deployments, secret access, or configuration changes that could impact system stability.
- Token Efficiency Auditing: Developers can see exactly where their budget went, with granular breakdowns of context usage by subagents, skills, and MCP servers.
- Privacy-First Architecture: The tool operates with a strict privacy model; the Nemotron-Mini-4B-Instruct model runs on the Space’s own GPU via ZeroGPU, ensuring that sensitive session data never leaves the environment or hits third-party APIs.
- Tool Intelligence Database: Her doesn't just list executed CLI commands; it identifies them using a database of top tools from Homebrew, npm, and PyPI, providing a one-line blurb for each.
What This Means
The release of Her marks a significant shift from "blind delegation" to "informed oversight" in the emerging AI agent era. As developers increasingly rely on tools like Anthropic's Claude Code to automate the development of entire features, the difficulty of auditing those actions has become a primary bottleneck for enterprise-wide adoption. By making the inner workings of a synthetic coder as legible as a human-authored pull request, Hugging Face is lowering the barrier to entry for high-stakes agentic automation. It moves the human role from "manual prompter" to "forensic reviewer," a necessary transition as the volume of AI-generated code continues to explode.
Technical Breakdown
The architecture of Her is designed to be both trustworthy and expressive, utilizing a split between a deterministic forensics engine and a generative prose layer.
- Deterministic Evaluation Engine: The core logic is built on a purely deterministic engine that parses JSON structures. This ensures that the numbers, token counts, and tool sequences are 100% accurate and do not shift based on model hallucinations.
- Local Language Model: It utilizes
Nemotron-Mini-4B-Instructfor the user-facing prose. This model is responsible for writing the English summaries and proposing suggestions based on best practices, but it never asserts raw data findings. - Interactive JOURNEY Graph: The frontend features a React-based interactive graph where every query is represented as a node. Nodes are sized by token cost, allowing developers to visually identify the "heaviest" parts of a session at a glance.
- Ask Her Copilot: A built-in assistant allows users to ask natural language questions about the trace, such as "Why did the agent reach for the production database?" The copilot cites specific turns and opens the exact tool call for inspection.
Industry Impact
For the broader AI industry, Her sets a new standard for transparency in autonomous workflows. Organizations that have previously hesitated to grant AI agents access to sensitive production environments now have a reliable forensic path to verify alignment and security. This move by Hugging Face reinforces the growing importance of the "observability layer" in the AI stack. It suggests that in the future, the most successful and trusted agentic ecosystems will not necessarily be the most powerful, but the ones that provide the clearest and most actionable audit trails for their human supervisors. We are seeing the birth of a new category of software: AI Forensics.
Looking Ahead
As agentic loops become more complex and cross-session memory becomes standard, tools like Her will likely evolve from passive forensic explorers into active, real-time steering and intervention platforms. The developer community is already demanding project-level views that can identify patterns and risks across weeks of automated development sessions. For now, Her provides the essential "black box" flight recorder that the autonomous coding industry has desperately needed to ensure safety and accountability in the age of synthetic intelligence.
Source: Hugging Face Blog(opens in a new tab) Published on ShtefAI blog by Shtef ⚡
