beiwocaishijia/gezhi-sheets

What it does

Gezhi Sheets is a local spreadsheet workbench for Windows. You select a range, describe what you need in Chinese, and the assistant produces a preview of extracted fields, converted text tables, formulas or cleaned data; nothing is written until you confirm. Every applied change can be undone once, and a preview is rejected if the underlying data changed after it was generated.

The six task areas are batch column filling (extract, classify, summarize, translate), converting free text such as orders or meeting notes into rows, data cleanup with duplicate flagging, table Q&A with grouped aggregates, formula generation with row-by-row trial calculation, and table operations such as sorting or highlighting. Recommendations for cleanup can be selected item by item, and suspected duplicates are only marked, never deleted automatically. Aggregates run as a restricted query plan executed by the program rather than by the model, and ambiguous operations trigger follow-up questions.

It suits people who work with customer feedback, product specs, order text or meeting notes in Excel and want AI help without uploading the whole workbook. The five supported providers are GLM, Kimi, Doubao, DeepSeek and Qwen, each configured separately with its own API key and model ID; the app does not switch providers automatically.

Getting started

Prerequisites: Windows and Node.js 22.13 or newer. The first release has no accounts or multi-user collaboration.

git clone https://github.com/beiwocaishijia/gezhi-sheets.git
cd gezhi-sheets
npm ci
npm run dev

Then open http://127.0.0.1:8001. In the app, open 模型设置 (Model Settings), pick a provider, enter its API key and model ID, and test the connection. Load the built-in feedback sample, import your own .xlsx or CSV, or paste directly from Excel. Select a range, type the instruction, generate the preview, review the original and suggested values, then click 应用 (Apply). After a first dependency install, the bundled 启动格知 AI.cmd launcher can be double-clicked instead.

For a production run: npm run build:app followed by npm start. The repository also ships component, AI-adapter and Playwright browser test commands; see the user and development guide and validation record.

Constraints

• Each sheet is limited to 1,000 rows and 100 columns, each workbook to 20 sheets, and Excel/CSV imports to 10 MB. The app targets sheets within 1,000 rows.
• Supported formulas in this first version are SUM, AVERAGE, MIN, MAX, COUNT, COUNTA, IF, ROUND, ABS, AND, OR, NOT plus basic arithmetic and same-sheet references. Macros, charts and complex layout are not guaranteed to round-trip; unsupported imported formulas fall back to the cached value with a warning.
• API keys are encrypted at rest with Windows DPAPI. Workbooks stay local, but the selected range is sent to the chosen provider when an AI task runs, and model fees and quotas are billed by that provider.
• The five adapters passed protocol-simulation tests; real account connections and output quality still depend on your own key and enabled models, and preset model IDs do not mean the account has access to them.
• Local data lives under %LOCALAPPDATA%\CE-Lab. Licensed under MIT.

HotRepo's take

Gezhi Sheets entered HotRepo's new-project list on 2026-11-13 for the JavaScript board, adding 5 stars over roughly 1.5 hours and finishing the sampled window at 105 stars and 33 forks. That is a single early snapshot, so it establishes the entry fact but not a trend.

The design choices visible in the repository are the more interesting signal: preview-and-confirm editing, one-step undo, blocked stale previews, and a program-executed query plan for aggregates all address the trust problem that keeps people from letting a model touch a real spreadsheet. The stated 144 passing tests cover component, backend and browser flows, and the README is explicit that screenshots use simulated AI responses rather than real model latency or accuracy.