CLI-Anything is an open-source framework from HKUDS (Hong Kong University of Data Science) that automatically generates production-grade command-line interfaces for any application with accessible source code, turning GUI-only software into agent-controllable tools. Released under Apache 2.0, the project has reached 38,600+ GitHub stars and 3,669 forks within two months, climbing the trending charts alongside a companion package, CLI-Hub, which acts as a centralized registry where agents can autonomously discover and install the generated CLIs. The slogan "Making ALL Software Agent-Native" captures the goal: instead of screen-scraping or brittle UI automation, agents drive real applications through structured CLIs that the framework built for them. ## The GUI Wall Most desktop and creative software, from Blender to GIMP to QGIS, exposes its capabilities through a GUI. AI agents have historically tried to bridge this gap with vision-language models clicking pixels or with hand-written wrappers around a handful of features, both of which are fragile and incomplete. CLI-Anything takes a different stance: if an application has source code, the framework can read it, understand the operations it supports, and synthesize a complete CLI that an agent can call with confidence. The result is what the project calls "authentic software integration" — the agent never touches the GUI, but the application still does all the real rendering. ## Seven-Phase Generation Pipeline Generating a usable CLI from arbitrary source is a non-trivial program-synthesis problem. CLI-Anything decomposes it into seven phases run by an LLM-driven agent. First comes source analysis, mapping GUI actions to internal APIs. Then command-architecture design produces a stable, ergonomic CLI surface. Implementation generates a Click-based Python CLI with persistent state. Test planning enumerates scenarios; testing actually invokes the real software in headless or visible mode to validate behavior. Documentation generation produces help text and examples, and the final phase publishes the package to PyPI. Each phase emits artifacts the next consumes, and refinement cycles can be triggered for incomplete coverage. ## CLI-Hub: A Registry for Agents Generating a CLI is only half the story. The framework ships with CLI-Hub, a centralized registry installed via `pip install cli-anything-hub` and used through `cli-hub install <name>`. The hub stores agent-discoverable metadata for every published CLI and includes a meta-skill that lets agents pick the right tool for a given task without human curation. This is the part of the system that most resembles a package manager for the agent era — instead of a developer browsing PyPI, an agent queries CLI-Hub at runtime to assemble the toolbox it needs. ## Dual Interaction Modes Each generated CLI exposes two interaction shapes. A stateful REPL mode keeps a persistent session, useful for interactive editing flows where an agent loads a document, runs a sequence of operations, and saves the result. A subcommand mode mirrors classic Unix tools, suitable for one-shot scripting and CI use. Outputs are always available in both human-readable text and structured JSON, the latter being what agents consume when chaining tool calls. ## Thirty-Plus Supported Applications The initial catalog covers 30+ applications spanning creative tooling (GIMP, Blender, Inkscape), productivity (LibreOffice, Zoom), media editing (Audacity, Kdenlive), scientific software (QGIS, FreeCAD), and specialized stacks (Godot, Ollama, ComfyUI). The diversity is the point: a single agent skill can now drive a 3D modeller, a raster editor, and a GIS workbench through the same calling convention, replacing what would have been three custom integrations. ## Installation and Limitations For Claude Code users, registering the marketplace and installing the plugin enables `/cli-anything <software-path>`, after which the seven-phase pipeline kicks off automatically. Other platforms (Pi, OpenCode, OpenClaw, Codex, GitHub Copilot CLI) have their own install paths. Reliable generation currently requires a frontier-class LLM such as Claude Sonnet, GPT-5, or Gemini 2.5 Pro; smaller models produce CLIs but with substantial gaps. Compiled binaries without accessible source degrade quality significantly, and complex applications may need iterative `/cli-anything:refine` runs to reach full surface coverage.