Top AI Product

The biggest bottleneck in AI agent development isn’t the models — it’s the software. An AI agent can write code, answer questions, and reason through complex problems, but ask it to batch-resize 500 images in GIMP or render a 3D scene in Blender, and it hits a wall. Professional desktop software was built for humans clicking buttons, not for agents issuing commands.

CLI-Anything, an open-source project from the University of Hong Kong’s Data Intelligence Lab (HKUDS), attacks this problem head-on. It’s a Claude Code plugin that analyzes any software’s codebase and automatically generates a production-ready command-line interface for it. No manual wrapping, no brittle screen-scraping, no toy reimplementations. The generated CLI calls the real application backend directly — LibreOffice generates actual PDFs, Blender renders real 3D scenes, Audacity processes audio through sox.

The project hit 7,200 GitHub stars in its first weeks of availability, landing on GitHub Trending in March 2026 and drawing attention from developers who have been frustrated by the gap between what AI agents can reason about and what they can actually do.

How the 7-Phase Pipeline Works

CLI-Anything doesn’t use templates or hand-written mappings. Instead, it runs a fully automated 7-stage pipeline that reads the target software’s source code and builds everything from scratch:

1. Analyze — scans the codebase, maps GUI actions to underlying API calls
2. Design — architects command groups and a state model
3. Implement — builds a Click-based CLI with REPL mode and JSON output
4. Plan Tests — generates a TEST.md with coverage strategy
5. Write Tests — implements unit and end-to-end test suites
6. Document — updates documentation automatically
7. Publish — creates setup.py and installs the CLI to PATH

The entire process runs with a single command: /cli-anything <path-or-repo>. Point it at a codebase, walk away, come back to a fully functional CLI with tests and documentation.

Every generated command supports a --json flag for machine-readable output, and standard --help flags let agents discover available commands at runtime. This self-describing design means agents don’t need custom integration code — they can explore capabilities the same way a developer would.

1,436 Tests, 9 Applications, 100% Pass Rate

The numbers behind CLI-Anything are what make it credible rather than just conceptually interesting. The team has generated CLIs for 9 major desktop applications, covering 1,436 test cases — 1,011 unit tests and 425 end-to-end tests — with a 100% pass rate across the board.

Here’s the breakdown by application:

These aren’t synthetic benchmarks. The e2e tests verify that the generated CLI actually drives the real application — Blender’s 58 end-to-end tests render actual 3D scenes, GIMP’s 43 e2e tests perform real image manipulations. That distinction matters because the whole point is to avoid the “works in testing, fails in production” trap that plagues screenshot-based automation.

Why Not Just Use RPA or Computer-Use Agents?

There are three main approaches to making AI agents control desktop software, and each has significant tradeoffs.

Screenshot-based agents (Computer Use, CUAs): These agents look at the screen, identify UI elements, and simulate mouse clicks. They’re flexible but inherently brittle — a UI redesign, a different screen resolution, or even a changed theme can break them. They’re also slow, since every action requires rendering, screenshotting, and visual analysis.

RPA (Robotic Process Automation): Traditional RPA tools like UiPath or Blue Prism record and replay UI interactions. They’re more structured than screenshot-based agents but still fundamentally tied to the GUI layer. When the interface changes, the automation breaks. And building RPA workflows for complex creative tools like Blender is impractical.

CLI-Anything’s approach: By generating CLIs that call application backends directly, CLI-Anything sidesteps the GUI entirely. The generated interfaces are deterministic — the same command produces the same result every time. They’re fast because there’s no rendering overhead. And they’re resilient to UI changes because they don’t depend on the UI at all.

The tradeoff is that CLI-Anything requires access to the source code, which limits it to open-source software. You won’t be using it to wrap Photoshop or Microsoft Office anytime soon. But for the massive ecosystem of open-source professional tools — and there are a lot of them — it’s a fundamentally more reliable approach.

The HKUDS Lab and the Bigger Picture

CLI-Anything comes from a prolific research group. The HKUDS Data Intelligence Lab, led by Dr. Chao Huang at the University of Hong Kong, has produced a string of popular open-source projects including LightRAG, RAG-Anything, AutoAgent, and AI-Researcher. Combined, their projects have accumulated over 77,000 GitHub stars.

The lab’s focus on “agent-native software” reflects a growing consensus in the AI development community: the next wave of productivity gains won’t come from smarter models alone, but from giving those models better tools. An AI agent that can reason about image editing is useful. An AI agent that can reason about image editing and directly execute GIMP commands to implement its reasoning is transformative.

The project is also actively evolving. As of March 12, 2026, the team submitted a PR for SKILL.md generation — a feature that would let generated CLIs describe their own capabilities in a format other AI agents and tools can understand. The roadmap includes expanding beyond the initial 9 applications to cover categories like AI/ML platforms (Stable Diffusion, ComfyUI), data tools (JupyterLab, Metabase), DevOps systems (Jenkins, Gitea), and enterprise software (GitLab, Grafana).

Getting Started

Installation takes two commands inside Claude Code:

/plugin marketplace add HKUDS/CLI-Anything
/plugin install cli-anything

Then point it at any software codebase:

/cli-anything <path-or-repo>

The generated CLI installs via standard pip install -e . and lands directly on PATH. While it’s built as a Claude Code plugin, the output is a standard Python package — meaning it works with any agent framework that can run shell commands, including Cursor, OpenCode, and others.

Frequently Asked Questions

What is CLI-Anything?
CLI-Anything is an open-source Claude Code plugin developed by the University of Hong Kong’s Data Intelligence Lab. It automatically generates production-ready CLI interfaces for any software with available source code, enabling AI agents to control applications like GIMP, Blender, and LibreOffice through structured commands rather than GUI automation.

Is CLI-Anything free?
Yes. CLI-Anything is fully open-source and free to use. It requires Claude Code to run the generation pipeline, which has its own usage costs through Anthropic’s API, but the plugin itself and all generated CLIs are free.

What software does CLI-Anything support?
It has been validated on 9 applications so far: GIMP, Blender, Inkscape, Audacity, LibreOffice, OBS Studio, Kdenlive, Shotcut, and Draw.io. The roadmap includes Stable Diffusion, ComfyUI, JupyterLab, Metabase, Jenkins, GitLab, and more. In theory, it can generate a CLI for any open-source software with a codebase.

How does CLI-Anything compare to RPA tools like UiPath?
RPA tools automate by recording and replaying GUI interactions, making them vulnerable to UI changes. CLI-Anything generates CLIs that call application backends directly, producing deterministic results without depending on the GUI. The tradeoff is that CLI-Anything requires source code access, limiting it to open-source software, while RPA can work with any application that has a visible interface.

Can I use CLI-Anything without Claude Code?
The generation pipeline runs inside Claude Code, so you need it for creating new CLIs. However, the generated CLIs are standard Python packages that work independently — any AI agent or script that can execute shell commands can use them.

You Might Also Like

• Heretic Just hit Github Trending and the ai World has Opinions
• Claude Code Security Just Dropped and it Already Found 500 Zero Days Nobody Knew About
• Pentagi Just hit 1 on Github Trending and Yeah its Worth the Hype
• Pageindex Just hit Github Trending and it Might Make you Rethink rag Entirely
• Claude Code Remote Control Just Turned my Phone Into a Coding Terminal and im Weirdly Into it