CLI-Anything teaches AI agents to control GIMP and Blender and gains 384 stars in a day
GIMP and Blender are designed for humans using a mouse, but AI agents work better with text commands. The CLI-Anything project lets an agent, such as Claude Code, read a program's source code and build a command-line interface (CLI) for it in 7 stages, complete with JSON responses and tests. In a single day, it gained 384 stars on GitHub. If an agent needs desktop software, generating such a wrapper is easier than writing scripts by hand.
AI-processed from GitHub Trending AI; edited by Hamidun News
Coding agents work confidently in the terminal: there are text commands, clear output and help pages. But most useful software is built differently. Blender, GIMP, LibreOffice, Zotero and Obsidian are designed for a person who clicks through menus.
For an agent to crop a hundred photos or assemble a 3D scene, developers usually have to write fragile scripts for each program. The CLI-Anything project from the HKUDS group aims to close this gap. A CLI (command-line interface) is simply a set of text commands for a program.
The idea is for an agent to build such a set for any software on its own. In one day, the repository gained 384 stars on GitHub Trending, meaning that many developers marked it as useful.
How CLI-Anything makes a program understandable to an agent
The scheme is simple. You give the agent a path to the program or a link to the repository, for example /cli-anything ./gimp. The agent then runs a pipeline of seven stages. First, it scans the source code and maps interface actions to program functions. Next, it designs command groups and the output format. After that, it builds a CLI on Click, with an interactive REPL shell, JSON output and action undo. Then it writes a test plan in a TEST.md file, implements unit and end-to-end tests, and records the results.
Imagine an agent needs to batch-process photos. Instead of "clicking" in GIMP, it calls commands for filters and export and receives a JSON response. There is no need to parse that response with regular expressions. Every command has a --help flag, so the agent discovers what is available by itself. The authors also stress that text commands match the format language models work with and are easy to chain together. The README also lists an optional fourth step: refining and improving the generated CLI.
CLI-Hub: ready-made wrappers
The second entry point into the project is simpler: instead of generating, you take a ready-made one. The CLI-Hub registry installs with pip install cli-anything-hub. Through it, you can search for, install, update and remove wrappers, and launch them with cli-hub launch <name>. For agents, there is a meta-skill: npx skills add HKUDS/CLI-Anything --skill cli-hub-meta-skill -g -y. After installing it, an agent with SKILL format support finds a suitable CLI for the task in the hub on its own.
The registry already contains many wrappers: Obsidian, Calibre, Zotero, draw.io, FreeCAD, tools for 3D meshes, video subtitles and GIS mapping. According to the changelog, some of them are covered by tests: the note-taking wrapper claims 134 tests, the screen recording editor 101, and one of the wrappers 258 commands in 17 groups. Supported agents include Claude Code, Codex, Cursor, OpenClaw, Nanobot, Pi, Reasonix and Antigravity.
Security and limitations
This approach gives an agent access to desktop programs, so security matters here. The changelog includes a fix for path injection in GIMP Script-Fu, protection for token files against escaping the directory and symbolic links, and URL validation and DOM sanitization in the browser wrapper. This is a good sign: the authors find and close such problems. But it also reminds us that every new wrapper adds to the attack surface.
There are other limitations. A wrapper does not replace the program itself: if a CLI relies on GIMP, Blender or LibreOffice, those must be installed separately. The quality of community wrappers will vary, because different people contribute them. A CLI generated by an agent should be read and tested before the agent is given access to working data. On Windows, Claude Code requires Git for Windows or WSL. The material contains no independent measurements of how much more reliable agents are with such CLIs than with other ways of controlling programs: the entire description is taken from the README.
What to do in practice
If you build agents for office, graphics or engineering software, first check the registry: the wrapper you need may already exist. If it doesn't, try the generator on a small internal utility and see how clean the JSON output and tests turn out. Evaluate the result not by GitHub stars but by the share of successful runs on your own tasks.
- Nov 21, 2026Anthropic to Unveil Claude for Factories: 17 Ready-Made Skills and 4 Integrations with Shop-Floor Software
- Nov 21, 2026AI Coding Assistant Under $10: Cursor at $8, Xiaomi MiMo at $6 a Month
- Nov 21, 2026Anthropic goes public and raises a record $107.3 billion
- Nov 21, 2026One Author Made an AI Comedy, 'Evacuator', in 2 Weeks and for 150,000 Rubles in Tokens
Key Questions Answered
How CLI-Anything makes a program understandable to an agent?
The scheme is simple. You give the agent a path to the program or a link to the repository, for example /cli-anything ./gimp. The agent then runs a pipeline of seven stages. First, it scans the source code and maps interface actions to program functions. Next, it designs command groups and the output format. After that, it builds a CLI on Click, with an interactive REPL shell, JSON output and action undo. Then it writes a test plan in a TEST.md file, implements unit and end-to-end tests, and records the results.
What are the key details on CLI-Hub: ready-made wrappers?
The second entry point into the project is simpler: instead of generating, you take a ready-made one. The CLI-Hub registry installs with pip install cli-anything-hub. Through it, you can search for, install, update and remove wrappers, and launch them with cli-hub launch <name>. For agents, there is a meta-skill: npx skills add HKUDS/CLI-Anything --skill cli-hub-meta-skill -g -y. After installing it, an agent with SKILL format support finds a suitable CLI for the task in the hub on its own.
What are the key details on Security and limitations?
This approach gives an agent access to desktop programs, so security matters here. The changelog includes a fix for path injection in GIMP Script-Fu, protection for token files against escaping the directory and symbolic links, and URL validation and DOM sanitization in the browser wrapper. This is a good sign: the authors find and close such problems. But it also reminds us that every new wrapper adds to the attack surface. There are other limitations.
Build & Ship Production AI Agents Without Boilerplate Code
Start your free 72-hour trial at Hamidun Academy (academy.hamidun.com) — interactive lessons on Claude Code & Cursor, live Tech Radar, and production agent templates.
The AI world, distilled — once a week
Seven stories that actually mattered, hand-picked. No noise, no reposts, no press releases.
Done! Check your inbox for a confirmation.