When an autonomous agent begins to hallucinate, execute dangerous commands, or burn through your token budget, you need a forensic audit, not another chat interface. The new open-source tool Her · हेर (Marathi for “detective”) is designed specifically to interrogate the logs left behind by Claude Code sessions. It transforms opaque, write-only JSON traces into a readable narrative, helping developers and makers understand exactly where their agents went wrong before they can deploy to production.
Turning raw logs into a readable case file
Every interaction with an AI coding agent generates a .jsonl file containing every turn, tool invocation, and token consumed. In practice, these logs are rarely examined. A developer might spend an hour trying to decipher thousands of lines of JSON to figure out why an agent made a risky move, where the context window budget was actually spent, or which subagent silently consumed half the execution time.
Her solves this by ingesting the session file and reconstructing the events in plain English. It flags high-risk activities—such as deployments, configuration changes, and secret exposure—and traces each action back to the specific turn where it occurred. The tool also breaks down token consumption, detailing which tools, subagents, skills, and Model Context Protocol (MCP) servers were utilised.
Crucially, the tool only offers suggestions when a named, fixable pattern is detected. These recommendations are grounded in Anthropic’s guidelines and community best practices. Her acts as a silent partner: it proposes improvements without asserting facts and remains quiet when there is nothing significant to report.
Interactive investigation and project-wide analysis
The interface includes a built-in copilot feature called “Ask Her”. Users can query the trace, such as asking “why was this tool used?”, and the system will answer from the log data, cite the relevant turns, and open the exact tool call for inspection. You can drop a single file for a session-level view or upload multiple files to build a project-wide view, allowing you to hunt for patterns or errors across many sessions simultaneously.
Privacy and deterministic evaluation
For security-conscious teams, the architecture is noteworthy. No third-party AI APIs are called. The underlying model, Nemotron-Mini-4B-Instruct, runs entirely on the Hugging Face Space’s own GPU infrastructure via ZeroGPU. Your session data is uploaded only to a private, auto-deleted namespace specific to your run, ensuring that no sensitive information leaves the server.
The tool’s reliability stems from a split evaluation engine that is purely deterministic. The AI model is used solely to generate the English prose and propose softer suggestions. It never asserts a finding as absolute fact, meaning the numerical metrics and forensic results do not fluctuate based on the model’s mood or temperature settings.
Offline tool identification
A useful detail often missing from other log viewers is the ability to identify external CLI tools. Her ships with a database of popular tools from Homebrew, npm, and PyPI. This allows the system to name most offline tools with a one-line description without needing internet access during the analysis. When actual deployment tools, database clients, or development servers are executed, Her flags this activity for the second look it deserves.
Origin story and technical stack
The project was built over a weekend. It began as an “operator’s view”—a journey graph where every query is a node sized by cost, with the heaviest node glowing red—created for a friend. Another friend requested a simpler interface, leading to the creation of an executive report view which is now the default. The initial lack of tool identification prompted the development of the offline database.
Technically, the frontend is a React application served directly from a Gradio server. The deterministic engine handles the forensic analysis, while the Nemotron model generates the narrative prose.
When your agent loses its mind, call Her. 😉
Key takeaways
- Turns logs into stories: Her converts write-only
.jsonltraces from Claude Code into plain English narratives, highlighting risky moves and token usage. - Privacy-first architecture: The tool runs locally on Hugging Face Spaces using
Nemotron-Mini-4B-Instructand ZeroGPU, with no third-party API calls and auto-deleted session data. - Deterministic reliability: The evaluation engine is purely deterministic, ensuring that findings and metrics remain consistent regardless of model updates.
- Offline tool detection: An included database of popular packages allows the tool to identify and name CLI tools from Homebrew, npm, and PyPI without requiring an internet connection.
Stay ahead of AI. Get the most important stories delivered to your inbox — no spam, no noise.
