# CLI-Anything: Turn Any Software Into a Tool Your AI Agent Can Use (55K Stars) > Source: > Published: 2026-10-19 11:31:09+00:00 AI agents are great at reasoning but terrible at using real software. CLI-Anything is the open-source fix with 55,000 GitHub stars: it generates CLI harnesses that turn any app into commands an agent can run. Blender, GIMP, LibreOffice, Zoom, FreeCAD, Obsidian, dozens more. It works with Claude Code, Cursor, Codex and other agents, and a hub command installs ready-made ones in seconds. Here is how it works, what the 18 real-world demos prove, and why it might replace GUI-based agents entirely. That line is the whole thesis of CLI-Anything, and it is why the repo at HKUDS/CLI-Anything picked up 55,000 GitHub stars in about seven months. AI agents can reason, plan, and write code, but ask one to actually use professional software, a 3D modeler, a video editor, a CAD tool, and it falls apart. The current workarounds are fragile screen-scraping automation, limited APIs, or watered-down reimplementations that miss most of what the software can do. CLI-Anything attacks the problem from a different angle. Instead of teaching the agent to click buttons like a human, it generates a command-line harness for the software, a real CLI with structured commands and JSON output that the agent drives directly. No screenshots, no mouse coordinates, no brittle UI automation. The agent talks to the actual software backend through a clean terminal interface. The project comes from the Data Intelligence Lab at the University of Hong Kong (HKUDS), the same group behind RAG-Anything, and it ships with a full tech report on arXiv plus 2,400+ passing tests. This is a research-backed tool, not a weekend hack. The core workflow is almost suspiciously simple. You point the CLI-Anything generator at a codebase, say a 3D application or a workflow tool, and run its seven-phase harness pipeline. It analyzes the architecture, designs the CLI, implements it, writes tests, generates docs, and can even publish the result, largely automatically. The README's own shorthand is: throw the codebase at /cli-anything and you get a full CLI your agents can use. The generated CLIs are not toys. They call the real application backends: LibreOffice generates actual PDFs, Blender renders real 3D scenes, Audacity processes audio. The README stresses this twice: zero compromises, zero toy implementations. Output comes in both JSON for agent consumption and human-readable formats for debugging, which solves the structured-data problem agents constantly hit. It works with the agents people actually use: Claude Code, Cursor, Codex, Pi, OpenClaw, and nanobot are all named as supported. If your agent can run shell commands, it can use a CLI-Anything harness. You do not have to generate anything yourself to get started. CLI-Hub is the project's registry of community-built CLIs, and installing one takes one command: ``` pip install cli-anything-hub cli-hub install ``` The hub lets you browse, search, install, update, and uninstall harnesses from one place, and there is a web frontend at clianything.cc to explore what exists. The registry keeps growing through community contributions: recent merges include harnesses for Obsidian, QGIS, Safari, n8n, Godot, Kdenlive, Calibre, and dozens of others. If you want to build your own harness and share it, there is a contributor signup flow, and merged contributors get recognized in the community. There is also a wishlist where you can request a CLI for specific software. The whole thing runs like an open-source app store for agent tooling, which is exactly what it is. CLI-Anything ships with 18 real-world demonstrations, and they are the reason the repo keeps getting shared. Agents use the generated CLIs to build a Mars rover in FreeCAD, design a drone in Blender, play a video game, generate subtitles, build CAD models, create 3D scenes and diagrams. These are end-to-end artifacts produced by agents driving real software, not scripted walkthroughs. The demos matter because they answer the only question that counts: does the agent actually get work done? A harness that technically exists but cannot complete a real task is worthless. The 18 demos, backed by 2,400+ tests with real software validation, are the project's evidence that the approach survives contact with reality. There is also a preview and trajectory-loop system that lets you watch what the agent is doing as it works, which is useful both for debugging and for the sheer fun of watching an agent model a drone. The README makes a bold claim: CLI-Anything can flat-out replace GUI-based agent approaches. The argument is hard to dismiss. GUI agents take screenshots, guess where to click, and break every time an interface updates. It is the most fragile possible way to automate software. A generated CLI gives the agent a deterministic, structured interface to the same capabilities, and once the software is agent-native, you can even synthesize agent tasks, evaluators, and benchmarks entirely through code and terminal, automated and iteratively refinable. The other angle is unification. Instead of juggling fragmented web service APIs, you can feed API docs or SDKs to the generator and get one stateful CLI wrapping them into coherent command groups. One tool instead of dozens of raw API calls, stronger capabilities, fewer tokens. The honest limitation: someone still has to generate or install a good harness for the software you care about, and niche or closed-source apps without accessible backends will be harder. But for open-source software, which the project targets first, the coverage is growing fast. The README advertises a five-minute quick start, and it is roughly true. If you just want to use existing harnesses, install the hub with pip install cli-anything-hub, browse the registry, and install what you need. You need Python 3.10 or newer. If you want to make your own software agent-ready, install the CLI-Anything plugin or skill for your agent, then run the seven-phase generator against the codebase. The docs live in the repo, the tech report is on arXiv for the full design, and the README is maintained in Chinese, Japanese, and German alongside English, which says something about how global the contributor base already is. At 55,000 stars, Apache 2.0 licensed, and actively merged multiple times a week, this is one of the most credible bets in the agent-tooling space right now. If you run AI agents and you have software they cannot touch, this is the repo to star.