Word GPT Plus: an Office add-in that lets an agent edit the document you have open
Word GPT Plus is a word add-in which integrates the AI&Agent into Microsoft Word.
At a glance
- What is it?
- Word GPT Plus is a Vue-based Microsoft Word task pane add-in that puts chat and a LangChain-driven agent next to your cursor, with OpenAI, Azure, Gemini, Ollama and Groq as model backends. The interesting part is not the chat box, it is that the agent can read and write the .docx you are editing.
- Who is it for?
- Adopt Word GPT Plus if your team already lives in Word, needs an OpenAI-compatible or local Ollama endpoint, and is comfortable sideloading a manifest per machine. Do not adopt it if you need tracked changes, .doc support, or a managed deployment across an organisation, because the README describes a manual sideload and states that only .docx files work.
- Can I use it commercially?
- Yes. MIT is a permissive licence: you can use, modify and sell software built on it, as long as you keep its copyright and licence notices.
- Is it still maintained?
- Yes. The repository received new commits within the last day.
- What is it written in?
- Mainly Vue, according to GitHub's language statistics.
Answers come from the project's GitHub data, last synced on September 15, 2026, and from our analysis. They are not legal advice.
DEEP OPEN-SOURCE ANALYSIS
The problem: model output that has to be pasted back into Word
Most writing assistants live in a browser tab. You copy a paragraph out of your document, paste it into a chat window, get a rewrite, then paste it back and fix the formatting by hand. Word GPT Plus removes that round trip by placing the model inside Word's task pane. The README describes the goal plainly: generate, translate, summarize and polish text "directly within your documents", without leaving the Word environment. The audience is narrow and identifiable. It is people who write long documents in Word and want translation, polishing, summarization or grammar correction applied to a selection rather than to a paragraph they retyped. The Quick Actions list in the README maps exactly onto that: one-click translation across 40+ languages, text polishing, academic writing enhancement, summarization, grammar checking. The second audience is narrower still, and it is the reason this project is worth a second look. Agent Mode, which the README says is powered by LangChain, is aimed at people who want the model to act on the document rather than describe what to change: insert text, create tables, manage bookmarks, search and replace. That is a different product from a chat sidebar, and it is where the real trade-offs sit.
How the agent actually touches the document
The architecture visible in the README is a task pane web app talking to model providers, with a tool layer in between for Agent Mode. The README lists the built-in Word tools by name: web search, insert text, format content, create tables, manage bookmarks, and search and replace. It also states that the agent can read, write and modify Word documents, that it does multi-step reasoning with conversation memory, and that responses stream with the thought process shown in collapsible details. That last detail matters more than it sounds. A multi-step agent that edits a live document is hard to trust if you cannot see the steps, and exposing the reasoning trace is the project's answer to that. The data flow is otherwise conventional for an Office add-in: the pane runs in a webview, holds provider credentials and settings in local storage, calls the chosen provider's HTTP API with streaming enabled, and applies the result through Word's document APIs. The README states that local storage is used for "complete privacy", which is accurate as far as it goes: prompts and keys are not routed through a project-operated server in the instant-use path. It does not mean the content stays on the machine, because the prompt still goes to whichever provider you configured.
Providers, models and where the requests go
The provider list is the most concrete part of the README. OpenAI (with a note that it is compatible with DeepSeek and other OpenAI-compatible interfaces), Azure OpenAI with custom deployment names, Google Gemini including Gemini 3 Pro/Flash, Gemini 2.5 Pro/Flash and AQA, Ollama for local deployments only, and Groq with Llama 3.3/4, Qwen3 and Kimi-K2 among others. Custom model names are supported for every platform, and the customization section mentions custom base URLs and proxies. The practical consequence is that this add-in is not tied to one vendor, and an Ollama setup keeps inference on the machine. Temperature and max tokens are configurable per provider, which is the right granularity: a translation preset and an agent preset want different settings. The README does not document how provider settings are stored beyond "local storage", nor whether keys are encrypted at rest. Treat that as unverified. If your threat model includes a shared Windows profile, assume the key is readable by anything running as that user until you check the code.
Getting it into Word: manifest, sideload, self-host
There are two installation paths. The instant-use path is a single file: download release/instant-use/manifest.xml, save it to a folder such as C:\Users\username\Documents\WordGPT, then sideload it following Microsoft's instructions for a network shared folder catalog. The README notes that users in China may need to add msq.pub to proxy rules or self-host instead, which tells you the instant-use manifest points at a hosted origin. The self-hosted path is for people who want their own backend. Docker is the shortest route: docker pull kuingsmile/word-gpt-plus followed by docker run -d -p 3000:80 kuingsmile/word-gpt-plus. Then you download release/self-hosted/manifest.xml and replace every occurrence of http://localhost:3000 with your server's address. Building from source needs Node.js 20+ and runs git clone, cd Word-GPT-Plus, yarn, yarn build, yarn run serve, again paired with the self-hosted manifest. There is also a Deploy to Tencent EdgeOne button that builds with npm run build and publishes ./dist. Software requirements are listed as Word 2016/2019 (retail), Word 2021 or Microsoft 365, plus the Edge WebView2 Runtime. The README states that only .docx files work and that the older .doc format is not compatible.
The limits: no tracked changes, no .doc, manual sideloading
The .docx-only constraint is a hard boundary and it is stated in the README, not implied. Anyone with a corpus of legacy .doc files has to convert first. The deployment model is the second constraint. Sideloading a manifest is a per-machine or per-catalog operation, and the README points at Microsoft's own guide rather than offering a central admin path. For a single writer that is fine. For an IT department asked to roll this out to two hundred people, the README gives you nothing. The third issue is the agent's write access. Tools that insert text, create tables and run search and replace operate on the open document. The README does not describe an undo boundary, a dry-run mode, or a diff before applying changes. Per-message actions are listed as replace, append or copy, which suggests the user confirms each response, but a multi-step agent issuing several tool calls in one turn is a different risk profile from a single replace. There is also no mention of Word's tracked changes or comments, so if your review process depends on change tracking, the agent's edits will not participate in it. Finally, the README does not describe telemetry, error handling when a provider rate-limits, or what happens to a partially applied agent run if you abort mid-stream. Abort is supported, but the document state after an abort is not documented.
What it is not: a comparison with a plain chat pane
The obvious alternative is the official assistant built into Microsoft 365 Copilot, or simply a browser tab with your model of choice. The difference in approach is where the model sits relative to the document. A browser tab has no document access at all: you supply the text and you apply the result. Copilot is integrated but tied to Microsoft's stack, licensing and data terms, and it is not something you can point at a local Ollama instance or a self-hosted gateway. Word GPT Plus sits between those two positions. It has document access through tools, and it lets you choose the backend, including a fully local one. The cost of that middle position is that you own the operational side: the manifest, the hosting if you self-host, the API keys, and the consequences of an agent that can write. If your requirement is a managed, supported, centrally administered assistant and you are already paying for Microsoft 365, the built-in option removes an entire category of maintenance work. If your requirement is a local model with document tools and no vendor dependency, this project is one of the few that offers it.
Maintenance, licence and what to check on upgrade
The project is MIT licensed, which permits commercial use, modification and redistribution provided the copyright notice and permission notice are retained. That is a permissive licence, and it is the reason self-hosting is viable for a company. It is not legal advice; if you redistribute the add-in internally or ship it to customers, have someone read the LICENSE file in the repository rather than this paragraph. On maintenance, the release history shows v2.0.0 in December 2025, v2.0.1 in January 2026, and a repository push in September 2026, with the README describing Agent Mode as LangChain-powered, which is the v2 line. The upgrade cost is concentrated in two places. First, the manifest: self-hosted installs embed your server address in manifest.xml, so if the upstream manifest changes shape you are diffing your edited copy against release/self-hosted/manifest.xml. Second, provider drift: model names are user-supplied, so a provider deprecating a model is a settings change rather than a code change, but the built-in model lists in the README will age. Pin the Docker image tag rather than tracking latest if you self-host, and keep your edited manifest under version control next to it.
Editorial conclusion
Adopt Word GPT Plus if your team already lives in Word, needs an OpenAI-compatible or local Ollama endpoint, and is comfortable sideloading a manifest per machine. Do not adopt it if you need tracked changes, .doc support, or a managed deployment across an organisation, because the README describes a manual sideload and states that only .docx files work. Before rollout, verify two things yourself: that the provider you intend to use is reachable from the machines that will run the pane, and that the agent's write tools behave the way you expect on a copy of a real document.
Community notes