CLI-Anything Reaches 54,400 Stars; GUI Agents Led CLI Agents 59.1% to 48.2% in One Study

HKUDS's CLI-Anything had 54.4k stars and 4.9k forks on its GitHub page on Oct. 18, and it sits on GitHub's trending list again. The star count says people want agents to stop clicking through desktop software. It says nothing about whether generated command-line wrappers do the job better, and the one comparison available to The Terminal suggests the answer depends on how much work went into the wrapper.

This is an analysis, not a test: The Terminal did not install CLI-Anything or run any generated CLI. It rests on the project's README and on a German-language analysis by remio.ai, whose numbers for the study below come from that one outlet.

What the README says it generates

The README's tagline is "Today's Software Serves Humans. Tomorrow's Users will be Agents." A plugin command such as /cli-anything ./gimp starts a seven-phase pipeline: analyze, design, implement, plan tests, write tests, document and publish. The output is a Python Click command-line tool with a REPL, --json output and a generated SKILL.md file that tells an agent how to call it.

A companion package manager, CLI-Hub, installs with pip install cli-anything-hub. The README lists Claude Code, Cursor, Pi, OpenClaw, OpenCode, Codex, Hermes, Reasonix, Qodercli and GitHub Copilot CLI as supported agents, with Goose marked experimental or community-maintained.

Two lines in the README set the standard. "Real software is a hard requirement — no fallbacks, no graceful degradation," it says, and "Tests fail (not skip) when backends are missing." The wrapper drives the actual GIMP or Blender, so the upstream app has to be installed. That makes the tool honest about what it controls and heavy to set up.

The test count does not agree with itself

The README's key-features section cites 2,280+ tests. Its own test table totals 2,461 in one place and 2,464 in another, a gap of up to 184 against the headline. remio.ai repeats the lower figure and says the harnesses cover 18 applications. None of the three numbers says how many tests exercise a real feature rather than a launch-and-exit path, and remio.ai makes the same point: the counts reflect the maintainers' chosen cases, not every application feature.

The comparison that matters

remio.ai summarises a study of 440 desktop tasks across 18 applications and 12 workflow categories, and treats it as the independent check on the CLI approach. The Terminal could not read the study itself, so these are remio.ai's figures.

1. Best CLI agent, original skills48.2 %
2. Best pure GUI agent59.1 %
3. Best CLI agent, verifier-guided skill expansion69.3 %

Source: remio.ai analysis of HKUDS CLI-Anything, accessed 2026-10-18

The strongest GUI agent beat the CLI agents that used the original skills by 10.9 percentage points. The CLI side overtook it only after verifier-guided skill expansion, ending 10.2 points ahead of the GUI agent. remio.ai reads that as evidence that missing coverage, not a flaw in the idea, accounted for much of the early gap.

For anyone choosing between the two, that is the practical finding. A generated CLI starts behind a GUI agent that can click anything, and it catches up only as someone extends the skills. A team adopting it is signing up for that extension work.

remio.ai's own conclusion is a hybrid: use commands where a wrapper covers the action, and fall back to GUI control for uncovered actions or visual checks. It also lists the failure modes it sees. Generated harnesses can be incomplete or faulty, especially when a weaker model writes them, and upstream application changes can break both the CLI route and the GUI route. Supply-chain trust and installation dependencies stay open questions in its account.

Release cadence and hardening

remio.ai dates three releases: v0.2.0 on March 30, v0.3.0 on April 24 and v0.4.0 on June 25, the last adding CLI-Matrix. A technical report by Yuhao Yang, Tianyu Fan and Chao Huang went to arXiv on June 2. The repository shows 893 commits on main.

The project's changelog lists security fixes that show what generated wrappers are exposed to: a GIMP Script-Fu path injection fix dated March 14, URL validation and DOM sanitisation for DomShell (#156), XML parsing routed through defusedxml (#296), and path-traversal and symlink hardening for Sketch token files (#304). A wrapper that shells out to desktop software inherits that software's attack surface and adds its own.

Version	Date (per remio.ai)	Note
v0.2.0	March 30
v0.3.0	April 24
v0.4.0	June 25	Adds CLI-Matrix

Stars are attention

remio.ai counted about 47,100 stars on Aug. 15, when the project ranked 12th on a trending list without a new release. The page showed 54.4k stars on Oct. 18, roughly 7,300 more in two months. Neither figure measures use. The README is Apache-2.0 licensed, and several harnesses are marked experimental, with ArcGIS Pro Windows-only and licensed.

The same caution applies to VoiceStudio's 73,300 stars and to the way a clever repository such as Esp32sim can draw attention before anyone measures it. On Windows, the README warns that Claude Code needs Git for Windows or WSL, or commands can fail with cygpath errors.

The number to watch is the next HKUDS release after v0.4.0. If skill expansion is where the study says the gains are, the changelog should say how much of it ships by default.