Her · हेर — 马拉地语中意为“侦探”。一个为你的 Claude Code 会话服务的侦探。
在此处试用：Hugging Face 上的 Her
每一次 Claude Code 会话都会留下痕迹——一个包含每一轮交互、工具调用和模型 token 的 .jsonl 文件。但在实践中，这个痕迹是只写不读的。几乎没人会去阅读 4000 行 JSON，只为了弄清楚智能体为何访问了生产环境、上下文预算究竟用在了哪里、或是哪个子智能体悄无声息地消耗了过半的运行资源。
Her 为你解读这一切。
原理很简单：将一个会话文件拖放到页面上，让她进行调查。她会用通俗易懂的英文重建所发生的一切，标记出高风险操作——部署、配置和生产环境变更、密钥——并将每一项追溯至其发生的具体轮次。
她会展示模型 token 的去向，使用了哪些工具、子智能体、技能和 MCP 服务器，并且——仅当存在一个可命名、可修复的模式时——基于 Anthropic 和社区的最佳实践，指出你本可以做得更好的地方。她提供建议，而非断言，并且在无话可说时保持沉默。
还有一个内置的副驾驶功能：向 Her 提问。问“为什么使用了这个工具？”，她会根据会话痕迹给出回答，引用相关轮次，并直接打开具体的工具调用。拖入一个文件可获得会话视图；拖入多个文件则可构建项目视图，并跨多个会话同时追踪一个问题。
全程不会调用任何第三方 AI API。模型——Nemotron-Mini-4B-Instruct——通过 ZeroGPU 在 Space 自身的 GPU 上运行。你的会话仅会上传至一个属于你本次运行的、自动删除的私有命名空间，且相关信息绝不会离开该环境。
使其值得信赖的关键在于分工：评估引擎是纯确定性的。模型仅用于撰写英文描述和提出较柔和的建议。它绝不会断言某个发现。当模型发生变化时，数据指标不会随之变动。
一个不错的细节：Her 不仅仅列出会话使用过的 CLI 工具——她还能识别它们。Space 内置了一个来自 Homebrew、npm 和 PyPI 的顶级工具数据库，因此大多数工具都能离线识别，并附带一行简介。当部署工具、数据库客户端或开发服务器被实际执行时，Her 会标记该活动，提醒你值得再次审视。
它在一个周末里成长起来。起初它只是一个操作员视角——一张旅程图，每个查询都是一个节点，节点大小由成本决定，成本最高的节点会发光——这是为一位朋友构建的。
我把它展示给另一位朋友，他想要更简洁的版本，于是这张图又衍生出一个高管报告，现在成了默认视图。然后第一位朋友问为什么他的 CLI 工具没有显示出来——工具数据库就这样诞生了。
前端是一个 React 应用，直接由 Gradio 服务器提供服务，确定性引擎负责取证分析，Nemotron 负责处理文本内容。
当 Claude 失去理智时，就召唤 Her。;)
在这里试用：Hugging Face 上的 Her
本文提到的 Spaces 1
Her · हेर — Marathi for “detective.” A detective for your Claude Code sessions.
Try it here: Her on Hugging Face
Every Claude Code session leaves a trace — a .jsonl file with every turn, tool call, and token. But in practice, that trace is write-only. Rarely anyone reads 4,000 lines of JSON to figure out why the agent reached for production, where the context budget actually went, or which subagent quietly burned half the run.
Her reads it for you.
The premise is simple: drop a session file onto the page and let her investigate. She reconstructs what happened in plain English, flags the risky moves — deploys, config and production changes, secrets — and traces each one back to the exact turn where it happened.
She shows where the tokens went, which tools, subagents, skills, and MCP servers were used, and — only when a named, fixable pattern fires — what you could have done better, grounded in Anthropic’s and the community’s best practices. She suggests, never asserts, and stays silent when there’s nothing worth saying.
There’s also a built-in copilot: Ask Her. Ask “why was this tool used?” and she answers from the trace, cites the turns, and opens the exact tool call. Drop one file for a session view; drop several to build a project view and hunt a question across many sessions at once.
No third-party AI API is ever called. The model — Nemotron-Mini-4B-Instruct — runs on the Space’s own GPU via ZeroGPU. Your session is uploaded only to a private, auto-deleted namespace that belongs to your run, and nothing about it leaves the box.
The split that makes this trustworthy: the evaluation engine is purely deterministic. The model is used only to write the English and propose softer suggestions. It never asserts a finding. The numbers don’t move when the model changes.
One nice detail: Her doesn’t just list the CLI tools a session used — she identifies them. A database of top tools from Homebrew, npm, and PyPI ships with the Space, so most tools are named offline with a one-line blurb. When deploy tools, database clients, or dev servers are actually executed, Her flags that activity for the second look it deserves.
It grew over a weekend. It started as an operator’s view — a journey graph where every query is a node sized by cost, the heaviest one glowing — built for a friend.
I showed it to another friend who wanted it simpler, so the graph grew an executive Report that’s now the default. Then the first friend asked why their CLI tool didn’t show up — which is how the tool database was born.
The frontend is a React app served straight off a Gradio server, with the deterministic engine doing the forensics and Nemotron handling the prose.
When Claude loses his mind, call Her. ;)
Try it here: Her on Hugging Face