What it is
CLI-Anything, from the Data Intelligence Lab at the University of Hong Kong (HKUDS), turns existing applications into tools an AI agent can use. Run inside an agent such as Claude Code, Codex or OpenCode, its seven-phase generator reads a program's codebase and builds a command-line "harness" for it, with JSON output, tests and a SKILL.md file that tells agents how to use it. The harnesses call the real software underneath, so the program itself must be installed.
CLI-Hub, installed with `pip install cli-anything-hub`, is a package manager for the harnesses the community has already built. They cover creative tools (GIMP, Blender, Kdenlive, Shotcut, MuseScore), office and knowledge tools (LibreOffice, Zotero, Obsidian), engineering software (FreeCAD, QGIS) and services such as Zoom and n8n. The README now reports 2,461 passing tests across about 30 harnesses.
It is licensed under Apache 2.0. The last tagged release, v0.4.0, came out in June 2026; CLI-Hub reached 0.4.1 on PyPI in July, and the last code change landed on 21 August 2026.
Our review
CLI-Anything is right about the interface and uneven on the delivery. Giving Blender or LibreOffice a command line with JSON output is faster and more predictable than computer use, where an agent clicks through screenshots, and CLI-Hub makes trying a ready-made harness a two-command job.
The weak point is that every harness is a separate piece of work. The README's own test table shows the spread: Blender has 208 tests including 58 end-to-end runs against the real program, while the Zoom harness has 22 unit tests and no end-to-end tests at all. The generator has the same dependency. Its limitations section says reliable output needs a frontier model such as Claude Opus 4.6 or GPT-5.4, that closed-source software without readable source code gives much worse results, and that one run rarely covers a whole application without further refine passes.
Security deserves the same care as any third-party code. The news log records a GIMP Script-Fu path-injection fix in March, path-traversal hardening for the Sketch harness and a switch to defusedxml for untrusted XML in May. A harness runs with your user's permissions, so reading it before an agent does is sensible.
The pace has dropped sharply. After near-daily merges from March to May, the news log stops on 30 May, the last code change was on 21 August and the September commits only touched the README, with 215 issues and pull requests open. The 2,461-test figure is impressive, but a benchmark of how often agents actually finish tasks with these harnesses is still an unticked item on the roadmap.
Where it earns its place
A 3D artist who scripts Blender renders
Have Claude Code batch-render scene variations overnight
Install the Blender harness from CLI-Hub and let the agent call it with JSON output instead of writing bpy scripts from scratch. Free apart from model usage; renders run on your own machine.
An internal-tools engineer at a mid-sized engineering firm
Give agents access to an in-house desktop application that has no API
Run the seven-phase generator on the application's source with a frontier model, then run refine passes until coverage is good enough. Expect several iterations and a code review before anyone else relies on it.
A research group manager using Zotero and LibreOffice
Automate reference management and document conversion without paying for an automation platform
Community harnesses cover both, at no licence cost. The risk is maintenance: a harness written by one contributor may not keep up with the next Zotero or LibreOffice release.
Before you commit
- Setup
- Python 3.10 or later; pip install cli-anything-hub, then cli-hub install <name>. The generator installs as a plugin or skill for Claude Code, Codex, Cursor, OpenCode and others.
- Integrations
- Harnesses expose a command line and JSON output plus a SKILL.md, so any agent that can run shell commands can use them; installable through npx skills as well.
- Data and privacy
- Everything runs locally against software you install; no hosted service is involved, and data handling depends on the wrapped program and the model provider.
- Limits
- Coverage and test depth vary by harness; generating new ones needs a frontier model and readable source code; activity has slowed since May 2026.
What's good
- Agents drive the real program, not screenshots
- Two commands to install a community harness
- Best harnesses test against the real software
- Apache 2.0 licence
What's not
- Test depth varies widely between harnesses
- Generator needs a frontier model to work well
- No code changes since August 2026
- Each wrapped program must be installed separately
Best for
People who want an AI agent to operate desktop applications such as Blender, GIMP or LibreOffice from the command line.
Alternatives
CLI-Anything: common questions
- Is CLI-Anything worth it?
- A sound idea, that agents work better through a command line than through screenshots, delivered as a catalogue whose quality depends on who wrote each harness. We rate it 3.8 out of 5.
- How much does CLI-Anything cost?
- Free and open source under Apache 2.0. The wrapped applications and the agent you run it in are separate.
- What are the downsides of CLI-Anything?
- Test depth varies widely between harnesses; generator needs a frontier model to work well; no code changes since August 2026; each wrapped program must be installed separately.
- Who is CLI-Anything best for?
- People who want an AI agent to operate desktop applications such as Blender, GIMP or LibreOffice from the command line.
- Is CLI-Anything safe for company data?
- Everything runs locally against software you install; no hosted service is involved, and data handling depends on the wrapped program and the model provider.
- What are the best CLI-Anything alternatives?
- The closest alternatives we have reviewed are Anthropic Skills and MCP Python SDK.
Answers checked Oct 22, 2026.