niuma

niuma is an AI agent that lives in your terminal. It does not just talk about your code - it drives a real browser, edits Office documents, and

generates or retouches images, all from one full-screen TUI.

WEB AUTOMATION, END TO END

Most AI tools can describe a bug in a web app. niuma can go and find it.

- Drives the Chrome you already use, with your existing logins (attaches over CDP - no bundled headless browser, no re-login)

- 14 built-in actions: open a page, read it, click, fill and submit forms in one batch, download files, take screenshots, run JS in the page

- Captures browser console errors, uncaught exceptions and failed requests (4xx/5xx) with the source URL and line number

- Inspects the request list and reads xhr/fetch response bodies, so "HTTP 200 but the business code says failure" becomes visible

- Intercepts and mocks requests on the fly (fulfill / abort / delay) to reproduce error and slow-network states

- Structured assertions (URL match, text present/absent, element visible, element count, no console error, request status) return pass/fail instead of

making you eyeball the page

- Waits on a selector or network idle, and still reports what happened while waiting

- Guardrails: never fills password / OTP / card fields; payment, order, delete, publish and other high-impact submits stop for confirmation

The loop: open -> fill -> catch the console or network error -> fix the code -> re-run -> assert. Turn a bug report into a verified fix without leaving

the terminal.

OFFICE FILES, READ AND WRITE

- Excel: read a summary (shape, column names, inferred types, non-null rate, preview), filter by column, aggregate (sum / mean / min / max / count /

unique), then edit cells, rows and columns or delete rows; CSV works too

- Word: read paragraphs, headings and tables; generate documents; edit paragraphs and table cells; local and global find-and-replace; add or remove

hyperlinks

- PowerPoint: read text, tables, images and chart placeholders; generate 16:9 decks; rewrite titles, bullets and tables; global text replacement; add

or delete slides; add hyperlinks

- PDF: read text, table of contents, summary and keyword search; generate PDFs; delete or rotate pages, merge files, stamp watermarks

- Every edit returns a before/after diff, and existing files are backed up (.bak) before being overwritten

IMAGE GENERATION AND RETOUCHING

- /image - text to image with any image provider you configure, controlled by --size, --seed, --n and --quality

- /image2image - image to image: keep the composition and swap the style, or keep the layout and redraw a region from a reference image; supports

source crop, reference crop, mask, feather and context padding

- /imgedit - offline editing: crop, resize, rotate, split. No tokens, no network

- Results are saved to images/YYYYMMDD/ and returned as local paths

ALSO IN THE BOX

- Code knowledge graph: symbol definitions, call chains and dependencies for 12 languages (Python, JavaScript, TypeScript, Go, Java, Rust, Ruby, PHP,

C, C++, C#, Lua), returning file:line instead of grep noise

- Persistent memory across sessions and up to 10 parallel contexts

- /loop - plan, execute, verify, self-correct with a 3-strike retry rule

- Parallel read-only sub-agents that research in isolated contexts

- Multi-provider routing: OpenAI-compatible by default, Anthropic protocol supported

- Web search with general, code-documentation and GitHub modes

PRIVACY

No background upload. No telemetry. No silent rewrites. Local-first - it can run fully offline with local models. Your code only leaves the machine in

the requests you send to the provider you configured.

Free, portable, no signup: download niuma.exe, double-click, start working. A free token quota is included.

----------------------------------------------------------------------------------------------------------

niuma 是一个跑在终端里的 AI agent。它不只是聊代码——它能操作真实浏览器、读写 Office 文档、生成与修图，全部在一个全屏 TUI 里完成。

一、Web 自动化全流程

多数 AI 工具只能描述网页里的 bug，niuma 能去把它找出来。

- 驱动你本机已登录的 Chrome（通过 CDP 附加，不内置无头浏览器、不用重复登录）

- 14 个动作：打开页面、读取页面、点击、批量填表并提交、下载文件、截图、在页面内执行 JS

- 采集浏览器控制台报错、未捕获异常、失败请求（4xx/5xx），带来源 URL 与行号

- 查看请求列表并读取 xhr/fetch 响应体，让"HTTP 200 但业务码是失败"这种问题无处可藏

- 实时拦截并 mock 请求（fulfill / abort / delay），复现错误态与弱网态

- 结构化断言（URL 匹配、文本存在/不存在、元素可见、元素数量、无控制台报错、请求状态）直接给出 pass/fail，不靠肉眼判断

- 等待元素出现或网络空闲，等待期间照常采集，把"等异步报错"变成可控动作

- 安全护栏：绝不代填密码 / 验证码 / 银行卡字段；付款、下单、删除、发布等高危提交必须先确认

闭环：打开 → 填表 → 抓到 console 或网络报错 → 改代码 → 重跑 → 断言。不离开终端，把一条 bug 报告变成已验证的修复。

二、Office 办公

- Excel：读取摘要（行列数、列名、推断类型、非空率、预览），按列过滤、聚合统计（求和/均值/极值/计数/去重），编辑单元格、行、列或删行；CSV 同样支持

- Word：读段落、标题、表格；生成文档；改段落与表格单元格；局部与全局查找替换；新增或移除超链接

- PowerPoint：读文本、表格、图片与图表占位；生成 16:9 演示文稿；改标题、要点、表格；全局替换；增删页；挂超链接

- PDF：读正文、目录、摘要、关键词搜索；生成 PDF；删页、旋转、合并、叠加水印

- 每次编辑都返回改动前后 diff，覆盖已有文件前自动备份 .bak

三、AI 绘图与修图

- /image 文生图：支持多家图片 provider，--size / --seed / --n / --quality 精细控制

- /image2image 图生图：保留构图换风格，或保留版式用参考图局部重绘；支持源图裁剪、参考图裁剪、遮罩、羽化、上下文填充

- /imgedit 本地修图：裁剪、缩放、旋转、拆分，离线执行，不消耗 token

- 结果按 images/YYYYMMDD/ 落盘并返回本地路径

四、还有这些

- 代码知识图谱：12 种语言（Python、JavaScript、TypeScript、Go、Java、Rust、Ruby、PHP、C、C++、C#、Lua）的符号定义、调用链、依赖关系，返回

文件:行号，替代 grep 大海捞针

- 跨会话持久记忆 + 最多 10 个并行上下文

- /loop 目标驱动编排：规划 → 执行 → 验证 → 失败自纠（3 次重试）

- 只读子代理并行调研，结论回传，不污染主上下文

- 多 provider 路由：默认 OpenAI 兼容协议，支持 Anthropic 协议

- 联网搜索：通用网页、代码文档、GitHub 三种模式

五、隐私承诺

无后台上传、无遥测、无静默改写、本地优先——可完全离线接本地模型运行。代码只在你主动请求自己配置的服务商时才离开本机。

免费、免安装、免注册：下载 niuma.exe，双击即用，自带免费额度。