Navigating the intricate world of AI agent development with Claude can be incredibly powerful, but understanding an agent’s inner workings often feels like peering into a black box. Each Claude Code session generates a detailed trace – a voluminous .jsonl file – logging every turn, tool call, and token. Yet, sifting through thousands of lines of raw JSON to pinpoint a specific issue is a monumental, often impossible, task for most developers.
This is where Her (हेर) steps in, acting as your dedicated detective for Claude Code sessions. Named after the Marathi word for “detective,” Her transforms those overwhelming log files into clear, actionable insights. Imagine finally understanding precisely where your context budget went, which subagent quietly consumed half your run, or why your agent made a crucial decision, all laid out in plain English.
Unraveling Your Claude Code Sessions
Her simplifies the complex by allowing you to simply drop your session files onto its intuitive interface. It then meticulously reconstructs the entire session, presenting a coherent narrative of events. No more deciphering cryptic JSON; Her provides a human-readable summary, making debugging and optimization significantly faster.
A key feature of Her is its ability to flag risky or significant actions within your Claude sessions. This includes any deploys, configuration changes, modifications to production settings, or interactions with sensitive secrets. Each flagged event is meticulously traced back to the exact turn where it occurred, giving you immediate context and accountability.
Beyond identifying risks, Her offers granular insights into resource utilization. It clearly shows you where your tokens were spent, which tools and subagents were invoked, and even which MCP servers were utilized during the session. This detailed breakdown is invaluable for optimizing performance and cost efficiency.
Her doesn’t just present data; it also offers constructive guidance. When a clear, fixable pattern emerges, grounded in Anthropic’s and the community’s best practices, Her suggests potential improvements. These are always presented as helpful suggestions, never as assertions, and it wisely remains silent when there’s nothing truly valuable to add.
Proactive Insights and Privacy You Can Trust
Her also features a powerful built-in copilot: “Ask Her.” This allows you to query the session trace directly, asking questions like “why was this tool used?” Her will then provide a precise answer, citing the relevant turns and even opening the exact tool call for your review. This interactive debugger streamlines the analysis process, helping you get answers instantly.
Whether you’re focused on a single interaction or a broader project, Her adapts to your needs. Drop a single file for an in-depth session view, or upload several files to construct a comprehensive project view. The project view enables you to hunt for specific patterns or answers across multiple sessions simultaneously, offering a holistic understanding of your agent’s behavior over time.
One of Her’s clever details is its sophisticated tool identification system. It doesn’t merely list the command-line tools used; it actually identifies them. Shipping with a comprehensive database of popular tools from Homebrew, npm, and PyPI, Her provides a brief description for most tools, flagging crucial activities like the execution of deploy tools, database clients, or development servers for closer inspection.
In an age where data privacy is paramount, Her sets a new standard for trust. No third-party AI APIs are ever called. The powerful Nemotron-Mini-4B-Instruct model runs entirely on the Space’s own GPU via ZeroGPU technology. Your session data is uploaded to a private, auto-deleted namespace unique to your run, ensuring that absolutely nothing about your sensitive information ever leaves the secure confines of the box.
This commitment to security is underpinned by a crucial architectural split: the evaluation engine responsible for all core findings is purely deterministic. The integrated model is utilized solely for generating human-readable English prose and formulating softer suggestions. This separation ensures that Her’s analytical findings and core metrics remain absolutely consistent and unbiased, regardless of any model changes.
The Weekend Project That Became Your AI Detective
Her’s journey began as a passion project, growing organically over a single weekend. It initially emerged as an operator’s view, conceptualized as a journey graph where each query was a node sized by cost, with the heaviest ones glowing brightly. This initial iteration was built to help a friend understand their agent’s operational footprint.
User feedback quickly shaped its evolution. Another friend requested a simpler interface, which led to the development of the executive Report, now the default view. Later, the request for better CLI tool visibility sparked the creation of its extensive tool database, continually enhancing Her’s capabilities.
Technologically, Her leverages a modern stack, combining a responsive React frontend served directly off a Gradio server. This powerful frontend works in concert with a robust deterministic engine that performs the forensic analysis, while the Nemotron model intelligently handles the English prose generation. It’s a testament to efficient design and focused development.
Ultimately, when your Claude agent’s behavior becomes opaque, and you need to shine a light on its actions, you know who to call. When Claude loses his mind, call Her.
Source: Hugging Face Blog
