Gezhi Sheets: open source AI spreadsheet assistant hits 300 stars
An MIT licensed, Windows-only spreadsheet workbench that previews every AI edit before it writes, and lets you pick from five Chinese model providers.
Gezhi Sheets, an open source AI spreadsheet assistant maintained by beiwocaishijia, had 300 stars and 102 forks on GitHub as of 2026-11-17, eight days after the repository was created on 2026-11-09. The project is MIT licensed, written mainly in JavaScript, and lets a user select a range of cells, describe what they want in Chinese, and inspect a preview before anything is written back.
The repository's own description frames the goal plainly: let the spreadsheet understand what you say. The project README lists six task types, five model vendors, a Windows setup path and a long list of hard limits.
Six jobs, one review step
The tool groups its work into batch filling a column, turning free text into a table, cleaning data, asking questions of a sheet, writing formulas and doing table operations such as sorting or highlighting rows. Every task follows the same sequence: pick a range, describe the requirement, generate a preview, confirm, apply. A single undo reverses the last apply, and a preview is rejected if the underlying data changed after it was generated.
Two design choices stand out. Duplicate detection only flags candidate rows and explains why; nothing is deleted automatically. And aggregation is not handed to the model. Filtering, grouping, counting, summing, averaging and sorting run through a restricted query plan that the program executes, with the model limited to interpreting the request. When a request is ambiguous, the tool asks a follow-up question instead of guessing.
Keys, vendors and what stays on disk
Five providers are wired in: GLM from Zhipu, Kimi from Moonshot, Doubao through Volcano Ark, DeepSeek and Qwen. Each keeps its own API key and model configuration. The settings screen supports searching presets, typing a model ID by hand and, where the vendor allows it, fetching the account's model list. The last selection is remembered, and the tool does not silently switch to another vendor.
API keys are encrypted on the machine with Windows DPAPI and survive a restart. Workbooks stay local. When an AI task runs, only the selected range is sent to the chosen vendor, which means the "local" label describes the workbench, not the model calls.
Setup requires Windows and Node.js 22.13 or newer. The README gives the clone, npm ci and npm run dev path, then a local address on port 8001, plus a .cmd launcher for after dependencies are installed. The first release has no account system and no multi-user collaboration.
Limits and open questions
Every sheet is capped at 1,000 rows and 100 columns, each workbook at 20 sheets, and imports at 10 MB. Formula support is a whitelist: SUM, AVERAGE, MIN, MAX, COUNT, COUNTA, IF, ROUND, ABS, AND, OR, NOT, basic operators and references to the current sheet. Macros, charts and complex formatting are not guaranteed to survive import, and unsupported imported formulas fall back to whatever cached value exists, with a warning.
The honesty in the README is worth noting. The five vendor adapters have passed protocol simulation tests only. Real account connections still need a key to verify, and a preset model does not mean the account has access to it. The demo screenshots use fixed simulated responses and say nothing about real speed or accuracy. The project reports 144 passing checks: 98 component regression tests, 37 backend and adapter tests, and 9 browser flow tests.
What the source leaves open: how well the preview step holds up on messy real data, whether the Windows-only DPAPI choice will be ported, and what the roadmap looks like beyond the first version.
A worked example from the README
Customer feedback arrives as free text, but what you need is a record you can follow up on. Select a few feedback rows, type a request to extract the requirement category and a one sentence summary, check the previewed columns against the originals, then apply. Output lands in new columns by default, which keeps the source text visible for comparison.
What this suggests for people building apps
The spreadsheet is the least interesting part. The pattern is the point: have the model propose, show the proposal next to the original data, let a human confirm, write once, and allow exactly one undo. Keep the deterministic work, especially arithmetic and aggregation, in code rather than in the model. If you are adding AI features to an app you are building, the guide to putting AI inside your app covers the same division of labour. Draftship is not affiliated with Gezhi Sheets.
Primary sources
- beiwocaishijia/gezhi-sheets: 格知 AI · Gezhi Sheets|让表格听懂你的话。支持 GLM、Kimi、豆包、DeepSeek、Qwen 的本地 AI 表格助手。 · GitHub trending AI repositories, Nov 17, 2026
This article was drafted with AI assistance from the primary sources listed below and checked automatically against them. An editor hasn't reviewed it yet.