CLI-Anything Has 62,119 Stars and a README With Three Different Test Totals
Software / analysis
CLI-Anything Has 62,119 Stars and a README With Three Different Test Totals
The HKUDS project turns desktop software into agent-callable CLIs; its own per-harness lines sum to 158 more tests than it reports.
HKUDS's CLI-Anything, a GitHub project that asks an AI coding agent to write a command-line wrapper around desktop software, has 62,119 stars. Its README reports its own test results three different ways, and the per-harness lines in one block do not add up to the total printed under them.
The repository at github.com/HKUDS/CLI-Anything was created on March 8, 2026 and last pushed on November 20, according to GitHub's API. It is licensed Apache-2.0, has 5,634 forks and shows 284 open issues and pull requests combined. The README promises that "Agents control any application without APIs, GUI, rebuilding or complex wrappers."
What CLI-Anything does
The project is a plugin for coding agents such as Claude Code, Cursor and Codex. Point it at a codebase and it runs a seven-phase pipeline: analyze, design, implement, plan tests, write tests, document and publish. The output is a Python CLI with a REPL, JSON output and undo and redo.
A companion package manager, CLI-Hub, installs finished harnesses with pip install cli-anything-hub followed by cli-hub install <name>. The README's own rule for them is "Zero Compromise Dependencies — Real software is a hard requirement — no fallbacks, no graceful degradation." A GIMP harness needs GIMP on the machine; it drives the real program and does not imitate it.
For someone batch-converting documents or rendering scenes through an agent, that design matters. A harness for LibreOffice or Blender gives the agent text commands and JSON back where a screenshot-clicking agent would get pixels. The cost is that every harness is another Python package you install and trust.
The three test totals in one README
The README's badge, its test table and its text each give a different figure. Here are the numbers as printed.
| Where in the README | Tests stated |
|---|---|
| Marketing table, "Prototype vs Production gap" | 2,280+ |
| Badge and test table total | 2,461 |
| Plain-text test summary block | 2,464 |
The 2,461 figure carries its own breakdown: "1,732 unit tests + 579 end-to-end tests + 19 Node.js tests." Those three add to 2,330, which is 131 short of the total beside them. I summed the 30 harness lines in the plain-text summary block as well, from gimp at 107 to joplin at 134, and they come to 2,622, which is 158 above the 2,464 printed as TOTAL.
- Marketing table (2,280+)2280 tests
- Unit + e2e + Node.js parts, summed2330 tests
- Badge and table total2461 tests
- Summary block TOTAL2464 tests
- 30 summary lines, summed2622 tests
Source: CLI-Anything README on GitHub, accessed 2026-11-22; the last two bars are sums computed from its own figures
The mismatch may be stale counters left over from fast-moving merges. The README does not say. What it does show is that the headline "100% pass rate" cannot be checked against a single agreed figure.
What the passing count leaves out
The summary block also records tests that did not run. Wavetone skipped 2 backend-gated end-to-end tests, Eth2-quickstart skipped 3, Unreal Insights skipped 9 and Joplin skipped 1 on Windows. A "100% pass rate" that excludes skips is a statement about the tests that ran.
Two harnesses show no end-to-end coverage at all. Zoom lists 22 unit and 0 end-to-end tests, and NotebookLM lists 21 unit and 0 end-to-end. For a harness that wraps a live service, a unit test cannot show that the command reaches the service.
The README's counts of supported software also differ. A badge reads "18 Apps" and the text says "18 major applications," while the summary block lists 30 harnesses.
The research behind it, and what it does not report
The README links an arXiv tech report, "CLI-Anything: Towards Agent-Native Computer Use", submitted June 2, 2026 by Yuhao Yang, Tianyu Fan and Chao Huang. Its abstract argues that "This GUI-centric paradigm fundamentally misaligns with agent capabilities" and that harnesses remove "the lossy visual-to-computational translation that plagues GUI agents."
The abstract reports no task success rates, benchmark results or counts of supported applications. The claim that CLI beats GUI for agents is an argument in the paper, not a measured comparison on the page we read.
The security question in the changelog
The README changelog records one hardening pass worth reading. On May 19, 2026, a "quality/security pass" routed untrusted XML, SVG, ODF, MLT, MusicXML and CSL input through defusedxml (#296), and the Zoom harness downloads recordings from direct URLs (#294). Each of these is a harness reading files or URLs an agent was told to open.
That is the same exposure our readers met in another trending agent repo, where a README advertised 681 patterns and cited 433. A star count of 62,119 says people looked. A harness that parses untrusted ODF on your behalf is code you should read first.
What to check before installing a harness
Read the harness folder for the one app you need, not the repository README. Look at its test line in the summary block, its end-to-end count and whether the app it needs is installed locally. A comparable test of that kind of claim, on a different agent project, is in our piece on OpenComputer's 3.3x Devin claim.
The README's newest dated changelog entry is May 30, 2026, though GitHub shows a push on November 20. Nothing in either promises a date for reconciling the totals.
Sources
More in Software
- 01Levien's Next-100 Rust GUIs Post Counts About 20 New Toolkits and Bets Against egui's Immediate ModeThe Linebender author credits AI rewrites for the surge; egui's own README concedes the trade-off he is pressing on.
- 02C++26 Is Reopened: The 13–9–12 Ballot, the Enforced Label That Answers It, and What Is Still OpenSutter's Búzios report says the committee kept contracts, added one paper, and will send C++26 to a second ballot with no published date.
- 03OpenAI Says Codex Cannot Run in Docker for iOS Builds, and Docker Launches It With Codex's Sandbox OffBolin's November 18 post lists Seatbelt, Bubblewrap and MXC as Codex's three sandboxes; Docker's page defaults to --dangerously-bypass-approvals-and-sandbox.
- 04Brazil's Receita Federal Posts Tax Calculator Under MIT With 16 Open IssuesThe agency published three repositories on Nov. 16 for the new consumption taxes, and outside developers had filed bug reports against the first within days.