CLI-Anything Has 55,252 Stars and Three Test Totals, and Its Per-Harness Lines Sum to 2,622

CLI-Anything, a GitHub project that generates command-line wrappers so coding agents can drive desktop software, gained 606 stars today on GitHub's trending list and stands at 55,252. The repository's README reports three different test totals, and the GitHub API shows its last push on September 22.

The project belongs to the HKUDS organisation on GitHub, which the page does not otherwise describe. It is licensed Apache-2.0 and was created on March 8. Its latest tagged release is v0.4.0 from June 25, and the repository shows 4,986 forks, 81 open issues and 120 open pull requests.

What the install asks of you

The README's headline is "CLI-Anything: Making ALL Software Agent-Native." You run /cli-anything <software-path-or-repo> inside an agent such as Claude Code, Cursor, Codex or OpenCode, and a pipeline of seven phases (Analyze, Design, Implement, Plan Tests, Write Tests, Document, Publish) produces a Click-based CLI with a REPL, JSON output and undo and redo. Each CLI ships a SKILL.md file so an agent can discover it.

The friction is in the fine print. Many of the generated CLIs wrap real desktop programs, and the README says to "install that upstream application too." On Windows, Claude Code users need Git for Windows or WSL, or commands may fail with "cygpath: command not found." Some Claude Code builds do not recognise /cli-anything at all and need a plugin reload or the legacy entry form.

The README also marks several harnesses experimental, among them NotebookLM, Goose, Codex and Hermes.

Three totals for one test suite

The README's headline claim is a 100% pass rate. It then attaches three different totals to it, and a fourth falls out of its own arithmetic.

Where in the README	Count given
Badge at the top	2,461 passing
Sentence under the harness table	2,461 (1,732 unit, 579 end-to-end, 19 Node.js)
"Test Summary" block, TOTAL line	2,464
Sum of the 30 per-harness lines in that block (our addition)	2,622
Footer and the "Prototype vs Production" row	2,280 and "2,280+"

The sentence that gives 2,461 also gives its parts, and 1,732 plus 579 plus 19 is 2,330. The 30 per-harness lines were summed by a short script run against the README's raw text, and they come to 2,622, which is 158 more than the block's own TOTAL.

The same row says tests were "Battle-tested across 18 major applications," and the footer says "18 professional software demos." The summary block lists 30 harnesses.

The skip claim

The README's design principles say "Tests fail (not skip) when backends are missing, ensuring authentic functionality." Its own summary block shows skips. Wavetone skips 2 backend-gated end-to-end tests, eth2-quickstart skips 3, unrealinsights skips 9, and joplin skips 1 on Windows, for 15 in total.

Those four harnesses are listed beside a "passed" mark and a 100% pass rate. The README does not say whether skipped tests count toward its totals.

Stars against activity

The star count is a trending signal, not adoption. The GitHub API shows the last push to the default branch on September 22, 28 days before today, while 606 people starred it today. The README lists proposed harnesses, including Hermes skill #320, ArcGIS Pro #318 and Joplin #300, as new entries, and the page does not say which are merged.

Earlier coverage of star-heavy repositories: Harness2, a Google Research repository with 10,700 stars and three commits, and VoiceStudio, with 75,300 stars and an AGPL app around a CC-BY-NC model.

What would settle it

The README links a Security policy, but the page excerpt carries no reporting process. The repository shows no release tag since v0.4.0 on June 25, so the next tag is the first chance to see whether the counts get reconciled.