0:03CLI-Anything comes from HKUDS — the University of Hong Kong's data-intelligence lab that also built LightRAG. The premise: most software is built for humans (buttons, menus, windows), and agents are bad at those but great at text. So CLI-Anything gives any app a command-line interface an agent can drive — with one command, from the source code.
Why a CLI
0:55Three current ways to let an agent use a desktop app, all flawed: GUI automation (screenshots + pixel clicks — slow, breaks when the UI changes), an API (many apps lack one, or it covers a sliver), and a rewrite (a toy Python copy that misses most features). A CLI fixes all three: agents already know terminals; it's text in, text out; every command has a help page; and here every command also has a --json flag so the output is clean JSON the agent reads without guessing.
CLI Hub
2:11Two ways to use it. First, CLI Hub — a package manager for CLIs other people already made: pip install cli-anything-hub, then cli-hub-list to browse the registry. The catalog is broad — Blender, GIMP, Inkscape, Audacity, OBS Studio, LibreOffice, Kdenlive, Shotcut, draw.io, Zoom, Ollama, ComfyUI, FreeCAD, QGIS, Obsidian, Zotero. One caveat: these CLIs call the real app, so the app has to be installed locally too.
The meta skill
3:10To hand it to an agent, the project ships a meta skill — a skill that teaches the agent how to find and install other tools. npx skills add hkuds/cli-anything (with the cli-hub-meta-skill flag), working with Claude Code, Codex, OpenClaw, Nanobot, Antigravity, and other skill-supporting agents. Then a single prompt — "find appropriate CLI software in CLI Hub and complete the task" — and the agent searches the hub, picks draw.io, installs it, reads its skill file, and builds an architecture diagram, exporting a PNG. "I did not tell it which tool to use. It found the tool by itself."
The 7-phase generator
4:14When the hub doesn't have your app — or you have an internal tool — use the generator. Add the plugin marketplace (/plugin marketplace add HKUDS/CLI-Anything), install it, then run /cli-anything with a path or GitHub URL. It runs a seven-phase pipeline: analyze (read the source, map each GUI action to the code that does it), design (command groups, state model, output formats), implement (the CLI in Python's Click library — every CLI gets a REPL mode, JSON output, and undo/redo), plan tests (a test.md plan), write tests, document (run the tests, write results back, and generate a skill.md so other agents discover the CLI), and publish (a setup.py that installs it on your PATH). Naming convention: cli-anything-<software>.
Authentic integration
6:47The project's key phrase is "authentic software integration" — the CLI doesn't replace the app, it calls it. The LibreOffice CLI produces a real PDF through LibreOffice in headless mode; the Blender CLI renders with real Blender. Running a CLI with no command drops into REPL mode, which shows the scene name and a * when there are unsaved changes, so the agent always knows the current state. 7:19And because one pass rarely covers a full app, there's /cli-anything refine — it compares what the software can do against what the CLI can do and fills the gaps (optionally focused, e.g. "more batch-processing and filters"), additively, without deleting prior work.
Tests & harness.md
7:57The harness.md playbook encodes the rules the agent follows every phase — and some are genuinely good lessons: the CLI must use the real software (never a substitute), and an exit code of zero doesn't prove an export worked, so tests check the real output files — magic bytes, file structure, pixels, audio levels, duration. The README reports 2,400+ passing tests across ~30 harnesses, each with unit tests (fake data) and end-to-end tests (real software) — and if the real software isn't installed, the tests fail rather than skip, because a skipped test can hide a broken CLI. Demos include an agent building a Curiosity Rover in FreeCAD step-by-step, a sci-fi drone in Blender, one playing Slay the Spire 2 through the CLI, and one driving a live ArcGIS Pro session.
Limitations
9:01Three, stated honestly: (1) it needs a frontier model (the README lists Claude Opus/Sonnet and GPT-5.4) — small local models produce incomplete or incorrect CLIs, and generation on a big codebase burns tokens; (2) it works best with source code — compiled binaries drop quality and coverage sharply; (3) one run is often not enough — expect to run refine a few times. And the standing warning: a CLI an agent made is still agent-made code — read the tests and check the commands before handing it important files.
Claims checked
The repo is real. HKUDS/CLI-Anything — ~57k★, Python, Apache-2.0, "Making ALL Software Agent-Native" (the video said 54k+ stars; it's grown). HKUDS is the University of Hong Kong's data-intelligence lab (LightRAG).
The 2,400+ passing tests across ~30 harnesses and the specific CLI catalog (Blender, GIMP, OBS, etc.) are the README's/video's claims, not independently verified here.
Auto-captions render the CLI as "cle anything" and mark the command flags as "{dash}" — the actual command is cli-anything and the flags are --json, -g, -y.