CLI-Anything turns any desktop software into an agent-ready tool

CLI-Anything is an open-source framework that automatically generates command-line interfaces for complex desktop applications, turning traditional software into native, controllable tools for AI agents.

AI agents excel at reasoning and generating code, but they notoriously struggle when tasked with operating real-world desktop software. Asking vision models to click around GUI interfaces often breaks the moment a window moves, while waiting for official APIs or rebuilding software from scratch is impractical.

What it does

CLI-Anything uses the command line as a universal protocol between agents and software. By running a fully automated pipeline on software source code, it generates clean, structured CLI harnesses that agents can invoke via standard terminal calls.

Instead of parsing screenshots and simulating mouse clicks, an agent can pass discrete flags to edit image layers, adjust audio settings, or export render passes headless.

The developers have already validated the pipeline across 8 major open-source applications, including GIMP and OBS Studio, passing 1,298 unit and end-to-end test cases.

The community has rapidly expanded the registry with harnesses for QGIS, Calibre, Obsidian, Joplin, Rekordbox, and Unreal Engine tooling. Any agent capable of terminal execution—such as Claude Code, Cursor, OpenClaw, or nanobot—can plug these CLIs in immediately.

How to run it

Everything is hosted on GitHub alongside a dedicated community registry called CLI-Hub.

You can get started by installing the hub client via pip install cli-anything-hub, then pulling down any supported harness with cli-hub install followed by the app name.

The project also supports direct skill installation via npx skills, letting you inject pre-packaged workflows into compatible agent platforms in one step. It even includes preview trajectory loops, allowing agents to inspect intermediate outputs before committing changes.

Why it matters

We think wrapping complex software into text-based CLIs is vastly more practical than betting purely on vision-based GUI automation. Terminal interfaces provide deterministic outputs, lower latency, and zero screen-resolution headaches for frontier language models.

For power users and developers running local agents, this unlocks precise automation inside production-grade tools without leaking proprietary assets to cloud services. The main catch is that the automated generator currently requires access to application source code or underlying scripting bindings, meaning closed-source proprietary software will still require custom reverse engineering or native hooks.