A Local AI Spreadsheet Assistant That Shows You the Preview Before It Writes a Single Cell
What This Thing Actually Does
Gezhi Sheets is an AI spreadsheet assistant that runs on your own machine, with a tagline that roughly means let the spreadsheet understand what you say. You select a block of data, describe what you want in Chinese, and it helps you extract, organize, and calculate. The crucial part is that it does not silently rewrite your table. It shows you a preview first, writes only after you confirm, and lets you undo in one click. It supports five domestic LLM providers, GLM, Kimi, Doubao, DeepSeek, and Qwen, each configured independently. The whole flow is deliberately broken into steps: select a range, describe the need, generate a preview, confirm and apply, then undo once, with a verification pass in the middle. In plain terms, it aims to be a spreadsheet partner whose AI does not move too fast, turning scattered information into usable data while keeping the steering wheel in your hands.
Why It Blew Up Recently
This repo appeared in early November and picked up steam quickly, and I think it hit two sweet spots. One is that the domestic LLM ecosystem has matured: GLM, Kimi, Doubao, DeepSeek, and Qwen can all be called reliably, so users in China no longer have to fuss with overseas accounts. The other is that it played the local-and-privacy card well. API keys are encrypted on-device with Windows DPAPI, workbooks stay on your machine, and when you launch an AI task only the range you selected is sent to the provider you chose, not the entire sheet uploaded wholesale. For office workers handling customer records or purchasing data, that sense of control is genuinely reassuring. The README also states plainly that 144 tests pass and all five adapters passed protocol simulation, which does a lot for credibility.
The Interesting Bits
Its capabilities are grouped into six categories: batch fill, text to table, data cleanup, spreadsheet Q and A, formula assistant, and sheet operations. A few concrete examples make it tangible. Select a column of customer messages, say extract the request category and a one-line summary, and it adds two columns filled in row by row. Paste an order in plain text, say turn this into customer, product, quantity, and unit price, and it produces structured records whose column names you can still edit in the preview. Say summarize purchase amounts by supplier and list the top five, and it runs a restricted query plan, handing filtering, grouping, counting, summing, averaging, and sorting to the program rather than letting the model fudge the math. The neat one is expressing calculations in Chinese: say quantity times unit price and it generates a formula and trial-computes each row, with Decimal used for the numeric work to avoid floating-point errors. Vague instructions trigger a follow-up question, suspected duplicates are flagged but never auto-deleted, and suggested cleanups can be picked item by item.
On the model side, each of the five providers has its own API key and model configuration, with searchable presets, manual model IDs, model-list retrieval, and a memory of your last choice that will not silently jump vendors. The stack is vanilla JavaScript with Webpack on the front end, Node.js, Express, native fetch, and Zod on the back end, with ExcelJS and Papa Parse handling files. Data handling includes debounced autosave, version checks, atomic replacement, a last-known-good backup, and retry on failure, plus read-only protection, structure-operation checks, batch progress, and cancellation. The limits are spelled out too: up to 1000 rows and 100 columns per sheet, 20 sheets per workbook, a 10 MB import cap, and a whitelist of formula functions such as SUM, AVERAGE, COUNT, IF, and ROUND.
Who It Is For
If you work in operations, customer service, or purchasing, or you wrestle daily with customer messages, orders, and meeting notes, this can save a lot of copy-paste. The first release is positioned as a local workbench, without an account system or multi-user collaboration, so it suits individuals and small teams using it on their own. It needs Windows plus Node.js 22.13 or newer, and data stays in %LOCALAPPDATA%\CE-Lab so returning users need not migrate anything. Worth noting: AI semantic results still need a human check against the original text, and you should keep an eye on provider fees and quotas.
Quick Start
Clone the repo, install dependencies, and run the dev server, then open http://127.0.0.1:8001. If you would rather not type commands, there is a batch file in the project you can double-click to launch it. Once inside, head to model settings to enter one provider's key and test the connection, then load the feedback sample or import your own xlsx or CSV, select a range, describe what you need, review the preview, and hit apply. That is the whole loop.
How It Compares
- Built-in AI in Excel or WPS: tied to one ecosystem, usually uploads the whole sheet to the cloud, and rarely offers this level of cell-by-cell preview and one-click undo before writing.
- Online AI spreadsheet tools: strong on collaboration and templates, but your data leaves the machine and domestic access or Chinese LLM support may be awkward.
- Gezhi Sheets: local execution, only the selected range leaves your device, your choice of five Chinese LLM providers, and preview-plus-undo that makes control the top priority.
- The catch: Windows only, capped at 1000 rows per sheet, limited formula and layout fidelity, and no multi-user collaboration, so heavy lifting still belongs to professional tools.