Open-source MCP servers let Claude and Cursor edit WPS documents, sheets and slides directly

AI agents get hands-on access to WPS

A hands-on guide published on dev.to shows how individual developers can connect WPS Office to AI assistants through the Model Context Protocol, turning Claude and Cursor from advisors into operators that edit documents, fill spreadsheet cells and restyle slide decks directly.

According to the post, Kingsoft's WPS 365 open platform now natively supports MCP, the connectivity standard Anthropic introduced that the author compares to a universal port for AI tools. The practical shift is one of access: instead of drafting text for you to copy and format, an agent can write into the open WPS window itself, applying a formula to a cell or assembling a presentation without a manual paste step.

Three open-source bridges

The guide compares three GitHub projects that expose WPS to MCP clients. wps-skills, the author's recommendation for first-time users, is a cross-platform Node.js server exposing 243 tools — 82 for spreadsheets and 112 for presentations — and runs on macOS, Windows and Linux. wps-mcp-server is a lightweight Python option built on Windows' COM interface with roughly 11 core tools, aimed at users who prefer a minimal setup. wps-editor-mcp takes a dual-mode approach, combining document parsing with COM control and format conversion, which suits jobs like turning a Word file into a presentation or processing files on machines without an office suite installed.

A four-step setup

Most of the walkthrough covers wps-skills, condensed into four steps. First, clone the repository and run the install and build commands inside its wps-office-mcp directory; Node.js 18 or newer and an installed copy of WPS Office are prerequisites. Second, register the server with the AI client — a mcp. file in the project root for Cursor, or Claude Code's settings file — pointing a node command at the absolute path of the built server file. The guide warns that relative paths and tilde shortcuts often fail to resolve.

Third, install the WPS add-in, which the author flags as the step where most people get stuck: macOS users run a shell script and Windows users a PowerShell script, after which WPS must be quit entirely, all processes included, and relaunched before a Claude assistant tab appears in the menu. Fourth, launch the server directly to verify it; a startup success message confirms the pipeline is live.

It drives the window, not the file

Under the hood, the post describes a three-layer design. A skills layer made of SKILL.md files helps the model decide which tool fits a request; the MCP server layer exposes the individual tools such as write_cell and add_animation; and an execution layer drives the application itself, through a wpsClient bridge on macOS and PowerShell plus COM on Windows.

The distinction matters. The project does not rewrite document files on disk, which the author notes are essentially zip archives prone to corruption when edited repeatedly. It manipulates the live WPS window instead, effectively clicking and typing on the user's behalf while the document updates visibly. The guide includes a short Python example that opens an MCP session and calls tools to create a document and write a value into a spreadsheet cell, and suggests that natural-language requests for a styled ten-page deck are within reach of the same tool set.

Extensions and failure modes

Two further capabilities are noted. The WPS Notes client can itself act as an MCP service, letting an agent archive its reasoning or meeting minutes directly into a notes library, and enterprise users on the WPS 365 open platform can attach private knowledge bases to build assistants scoped to company data.

The troubleshooting section covers the common failures: the target document must already be open in WPS, the add-in installer may need administrator rights, module-not-found errors usually trace back to non-absolute paths in the config, and Windows COM errors point to checking that WPS is fully installed with COM automation enabled.

Why it matters

The long-standing gap in AI office tooling was never intelligence but write access — models could draft content but could not apply a format, enter a formula or adjust a slide. By pairing a free office suite with open-source MCP servers, individual developers get that access without waiting on an official API, and the setup cost is a config file plus a build step. WPS embracing MCP also makes it a first-class target for agent workflows, a notable move for a suite competing outside the Microsoft ecosystem. The trade-off is built into the architecture: an agent that drives your live editor has direct, unsupervised write access to working files, so the convenience arrives with a trust question attached.