When letting AI Agent process Office files, the difficulty is usually not “whether it can write a paragraph of text”, but whether it can retain the existing template, correctly update formulas and charts, see the typesetting results, and confirm that the file has been truly saved before delivery. OfficeCLI is a local command line tool for Agent, supporting the reading, creation, editing and rendering of .docx, .xlsx and .pptx without the need to install Microsoft Office.
It’s suitable for placement in local automation, CI document generation, or controlled Agent workflows. It should not directly obtain unlimited write permissions on shared disks, customer originals, or financial master files; it should start with copies and templates, and then deliver them after previewing and manual review.
What tasks is OfficeCLI suitable for?
OfficeCLI provides structured commands to operate Word, Excel, and PowerPoint. Common uses include:
- Read text, styles, tables, formulas or slide structures from existing documents;
- Generate reports, presentations and workbooks using JSON or commands;
- Batch search and replace, unify styles or fill data according to templates;
- Generate fixed-format weekly reports, quotes, and project summaries from a database or API;
- Render Office files to HTML or PNG and let AI and humans check the layout;
- Generate documentation in a CI or container environment without relying on desktop Office.
For one-time personal documents, a desktop editor may be faster; for repeated generation, batch processing, or scenarios that require Agent participation, command lines and structured output are easier to review, reproduce, and test.
Determine file boundaries before installation
Before any installation and automation, set these three boundaries:
- Input directory: The Agent is only allowed to read the templates or copies required for the task, and does not directly read the entire shared disk.
- Output directory: Write the generated results to a separate
out/or temporary directory to avoid overwriting the original files. - Delivery Actions: Uploading, emailing, overwriting shared files, and modifying financial data should all be manually confirmed.
Office documents may contain comments, revision history, hidden worksheets, external links, macros, or sensitive metadata. Before automating, you should confirm that the source is trustworthy and test with a copy. Don’t mistake “can read the file” for “can safely expose the entire contents of the file.”
How to install and confirm version
OfficeCLI provides single binaries for each platform and also supports package manager installation. Windows users can preferentially use configured package channels, such as:
|
|
You can also download the binary files for the corresponding platform from the project Releases. After installation, confirm the version first:
|
|
If you use download scripts or global npm installation, you should first check the repository, version, and enterprise software policy; do not put remote scripts from unknown sources directly into high-privilege terminals. For team environments, it is recommended to pin the tested version rather than letting each machine automatically track the latest version.
Create and preview PowerPoint for the first time
The minimal workflow is “Create → Write → Preview → Save”. For example, create a presentation and add a title page:
|
|
outline Ideal for quickly checking slide and text structure, html produces renderings that can be viewed in a browser. Don’t assume a file is deliverable based solely on command success: title overflows, overlapping shapes, font substitutions, and chart scaling issues are often not visible until the rendering results.
When you need to make continuous adjustments, you can start a real-time preview:
|
|
In the official example, the preview service will be opened on the local machine http://localhost:26315 by default. Only use this port in a local controlled environment, and do not expose the preview service to the public network without verification.
Let the AI Agent work according to “read, modify, view, and verify”
Directly asking the Agent to “generate a PPT” can easily result in uncontrollable rewriting of the entire file. A more stable way to prompt is to break up the actions:
- First read the structure of the target file or export JSON without modifying any content.
- Let the Agent list the page, cell, or paragraph paths that are scheduled to be modified.
- Change only one logical unit at a time, such as a slide or a worksheet area.
- Render HTML or screenshots and check content, overflow, alignment and data correctness.
- Save to the output directory, and finally manually review and decide whether to replace the original.
Examples of task constraints that can be used directly:
Read
templates/monthly-report.pptx, do not modify the original file. Generatesout/monthly-report-review.pptxbased ondata/summary.json; only replaces the title, three KPIs, and chart data. When finished, export an HTML preview listing all modified paths, unpopulated fields, and content requiring manual confirmation.
This is easier to review than “update monthly reports based on data”, and also prevents Agents from changing irrelevant masters, notes, or hidden content.
Differences between Word, Excel, and PowerPoint
| File Types | Common Things Agents Do | Key Checks Before Delivery |
|---|---|---|
Word .docx |
Replace paragraphs, tables, headers and footers, styles and table of contents | Page numbers, pagination, references, revision history, fonts and table breaks |
Excel .xlsx |
Write cells, formulas, data validation, charts and pivot tables | Formula results, reference ranges, hidden tables, filter conditions, chart data sources |
PowerPoint .pptx |
New slides, shapes, pictures, charts, notes and themes | Title overflow, element overlap, color matching, proportion, and show effects |
You need to be particularly careful with Excel: just because the formula can be written correctly does not mean that the business caliber is correct. The metric definition, date range, currency, unit, and rounding rules should be written in the task input, and let the Agent output the calculation source and outliers instead of just the final number.
Use template merging instead of duplicate generation
Repetitive reports should not require Agent to start from scratch each time. OfficeCLI’s merge can replace the {{key}} placeholder in .docx, .xlsx, or .pptx with JSON data:
|
|
The suitable approach is to manually determine the template and fields for the first time, and then the program or Agent will only prepare the verified data and perform the merge. This is more stable than rearranging each time, and can also separate “content generation” and “layout control”.
When a field is missing, have the process stop and report the missing item rather than continuing with a guess, an empty string, or data from the last customer. This is especially true for quotes, contracts, invoices, performance sheets and external-facing reports.
Pay attention to the timing of saving when editing multi-steps
OfficeCLI supports resident mode, which keeps documents open in memory to speed up multi-step operations:
|
|
In resident mode, the tool’s own read command can see the latest memory status; but other programs, synchronization clients, or upload processes may not be able to see it. Before handing it over to Word, Excel, renderer, Python script or network disk, execute officecli save or officecli close to ensure that the modifications have been written to the disk.
Batch operations also require a clear failure strategy. If each command is allowed to continue, it may produce “partially updated” files; it is more suitable to use --stop-on-error for critical documents and leave failed tasks to manual processing.
How to do quality check after generation
It is recommended to divide quality inspection into three levels:
- Structure Check: Confirm the number of pages, worksheets, paragraphs, placeholders, table rows and output paths.
- Data Check: Spot-check formulas, summary amounts, dates, chart ranges, and source data versions.
- Visual inspection: View HTML or PNG to check text truncation, margins, overlap, color and template consistency.
For batch-generated files, random sampling can be manually reviewed, and the input data version, template version, OfficeCLI version, command log and generation time can be retained. In this way, when an error occurs, it can be located whether the problem is with the data, template, command or model instruction.
FAQ
The file is generated but the style is wrong
First execute officecli view <文件> html or take a screenshot to preview, don’t just check whether the file can be opened. Confirm whether the Agent has changed the wrong object path, whether the theme or master has been overwritten, and whether the input data exceeds the text capacity of the original template. Write large text limits, image sizes, and number of charts into task constraints.
Excel data is correct but chart or formula is wrong
Check chart reference ranges, worksheet names, absolute/relative references, and date formats. For workbooks that contain macros, external data connections, complex plug-ins, or enterprise templates, verify compatibility on a copy first; automated tools should not bypass your organization’s security policy for macros and external links.
Agent overwrites the original file
Immediately stop subsequent writes and restore from version control, backup, or original template. The input path is then set to a read-only copy, the output path is fixed to a separate directory, and the Agent is required to report the target file before each write. Do not rely on the “please do not overwrite” prompt to protect files.
Summary
OfficeCLI allows AI Agents to process Word, Excel and PowerPoint in a local, non-Office installation environment, but the reliability comes from the process rather than a single command: limiting the input and output directories, starting from templates and copies, looping through read, modify, and verify, saving and then handing it over to external programs, and manually reviewing the data and layout. Think of it as a reviewable document automation component rather than an automatic publisher to gain both efficiency and control.
Official warehouse: https://github.com/iOfficeAI/OfficeCLI