TL;DR and context #
A free GitHub tool lets Claude Code run desktop apps without connectors. Where it beats MCP, and where connectors still win.
- CLI-Anything turns desktop software into terminal commands that Claude Code and other coding agents can run, with no MCP server needed.
- It does not click around the screen, instead it builds a tested command line tool on top of each app’s own files, API or headless mode.
- Ready made tools exist for draw.io, Blender, GIMP, OBS Studio, Zoom, LibreOffice and dozens more, installed through CLI-Hub.
- Use it for desktop and open source apps on your own machine, and keep MCP connectors for shared business systems that need sign in, permissions and logs.
A post going round this week says a GitHub repository “just killed MCPs”. Its author asked Claude Code to draw the system design for a food delivery app in draw.io and got a finished diagram, 11 boxes and 16 arrows, without opening the app. The repository is CLI-Anything, an open source MCP alternative for desktop apps with just under 60,000 GitHub stars, and it does let Claude Code work with software that has no MCP connector. MCP is still very much alive, though, because the two solve different problems, and picking the wrong one costs you either setup time or control.
CLI-Anything is an open source Python project from the Data Intelligence Lab at the University of Hong Kong (HKUDS), which says it wants to make all software “agent native”. Coding agents such as Claude Code work by running commands in a terminal and reading the result, while most desktop apps were built for a person with a mouse. CLI-Anything gives each app a command line of its own that an agent can use.
Connecting Claude to another app usually needs an MCP server. Popular business tools have one, but Inkscape, Audacity, Krita or a niche lab program usually do not. The repository now covers more than 90 applications, from Blender, GIMP and Kdenlive to LibreOffice, Zotero, Obsidian, n8n and Ollama.
How CLI-Anything works #
CLI-Anything does not watch the screen or simulate clicks, instead it builds a proper command line tool for each app, once, and that tool is what Claude uses. A coding agent builds it by following a written method in a file called HARNESS.md, in seven phases.
- Analyse: read the app’s source code and map what each button does to the function underneath.
- Design: plan the commands, the state they track and the output format.
- Implement: build the CLI, with JSON output and undo and redo.
- Plan tests: write a test plan.
- Write tests: build and run unit and end to end tests.
- Document: record the results and write a SKILL.md file so other agents can discover the tool.
- Publish: package it so it installs as a normal command.
Every tool returns structured JSON, so Claude never has to parse text meant for people, and every tool drives the real software rather than copying it, using LibreOffice’s headless mode for document conversion and Blender’s background mode for rendering.
That explains the viral demo. A draw.io diagram is a plain XML file in which each box is a shape and each arrow is a connection between two shapes, and the draw.io tool writes that file directly. The app is only needed for a PNG, PDF or SVG export, and even then it runs without a window, so the agent never had to open draw.io.
Other tools work differently. The Zoom tool calls Zoom’s official REST API, and you have to register your own OAuth app with Zoom first, so “connects to Zoom” means what it usually means: an API Zoom already provides.
How to use it with Claude Code #
1. Install the plugin inside a Claude Code session.
/plugin marketplace add HKUDS/CLI-Anything
/plugin install cli-anything
2. Install a ready made tool from CLI-Hub, the project’s registry, which is where most people should start.
pip install cli-anything-hub
cli-hub search drawio
cli-hub install drawio
cli-hub list shows everything available. If the app you need is missing, point Claude Code at a local copy of its source with /cli-anything ./<app>, and widen an existing tool with the refine command, for example /cli-anything:refine ./gimp "batch processing and filters". Refinement is incremental, so broad coverage takes several runs.
3. Install the app itself. The GIMP tool does nothing without GIMP, and draw.io exports need the draw.io desktop app.
4. Ask Claude in plain language, for example for “a system design diagram for a food delivery app in draw.io”. Claude then runs commands like these, taken from the tool’s documentation.
cli-anything-drawio project new --preset letter -o diagram.drawio
cli-anything-drawio --project diagram.drawio shape add rectangle --label "Order service"
cli-anything-drawio --project diagram.drawio shape add cylinder --label "Orders DB" --x 300 --y 100
cli-anything-drawio --project diagram.drawio connect add <source_id> <target_id>
cli-anything-drawio --project diagram.drawio export render diagram.png -f png
5. Open the file and check it. The project’s own method warns never to assume an export worked just because nothing failed.
Windows users need Git for Windows or WSL for the Claude Code plugin, and integrations for other agents such as Codex and OpenCode are labelled experimental or community maintained.
CLI-Anything vs MCP #
MCP, the Model Context Protocol, is the open standard for connecting AI apps to outside tools. In our explainer on how a company builds an internal AI assistant we described it as a common plug shape, so each tool does not need its own cable, and CLI-Anything is a different kind of plug.
| CLI-Anything | MCP server | |
|---|---|---|
| Best for | Desktop and open source apps on your own machine | Shared business systems such as CRM, ticketing and payments |
| Where it runs | Locally, with your files and permissions | Locally or hosted in front of a company system |
| Who builds it | A coding agent, from the app’s source or API | Usually the vendor or a developer |
| Context cost | Nothing until a command runs | Tool descriptions load into the conversation |
| Permissions and logs | Whatever your account can do, with no central log | Can enforce its own sign in and log every call |
| Works with | Any agent with a terminal | Any MCP client, including Claude, ChatGPT and IDEs |
The README argues that commands are structured, composable, self describing and deterministic, so one can be piped into another, scripted and run the same way by any agent. MCP servers have drawn complaints for filling the model’s context with tool descriptions, a problem Cloudflare tackled in the approach covered in our Cloudflare MCP explainer.
The repository also uses MCP itself. Its registry includes an ArcGIS Pro tool that drives a live session of Esri’s commercial mapping software over MCP, and a Safari tool built on a Safari MCP server, because when an app is closed source and only reachable while running, MCP is often the practical way in.
If the app lives on your computer and works with files, try CLI-Anything first. If it lives on a server and holds customer data, money or records someone may audit, use MCP or the vendor’s own connector, and remember that installing a connector is not the same as connecting it, as we found with Claude’s Marketing plugin.
CLI-Anything has limits of its own. The build starts by reading source code, so proprietary apps are reachable only if the vendor offers an API, scripting or a headless mode. Quality thins toward the long tail: the README badges cite 2,461 passing tests, while its feature table cites more than 2,280 across 18 major applications.
Installing a community tool also means running someone else’s Python with your permissions, and an agent that can run it can overwrite your files, which is what went wrong in our report on an AI agent that rebooted a server. Read what you install and pin versions. A tool on one employee’s laptop also keeps no central record of what the agent did, a weak position for anything involving personal data under the GDPR’s accountability principle and the transparency duties the EU AI Act places on deployers.
Browse the GitHub repository hub, see what marketing teams need to self-host five GitHub tools, or read how WebMCP gives AI agents a menu on websites.
Methodology and sources #
Industry Contents did not run CLI-Anything for this explainer. We read the public repository on 22 November 2026, including the README, the HARNESS.md method and the draw.io and Zoom tool folders, and took the draw.io commands from that tool’s own documentation. Star counts come from Trendshift and AlternativeTo and change daily. The food delivery diagram is the original poster’s and could not be independently verified. Sources: HKUDS/CLI-Anything, CLI-Hub, cli-anything-hub on PyPI, Model Context Protocol.
Frequently asked questions #
Does CLI-Anything replace MCP? #
No. It covers desktop and open source apps that have no MCP server. MCP remains the better route for shared business systems that need sign in, permissions and logs, and some CLI-Anything tools run over MCP themselves.
Is CLI-Anything free? #
Yes, the repository is published under the Apache 2.0 licence. You still need the target app installed, and some tools, such as Zoom, need your own account and API credentials.
Does CLI-Anything work with tools other than Claude Code? #
Yes. Any agent with a terminal can run the generated commands, and the project ships integrations for Codex, OpenCode and others, though those are labelled experimental or community maintained.
Can CLI-Anything control any software? #
Not quite. It works best on open source apps, because the build reads source code. Closed source software is reachable only through an API, scripting layer or headless mode the vendor provides.