CLI-Anything

Operator's take

The thesis is the interesting part. Most attempts to make existing software "agent-friendly" go one of two ways: a brittle browser-automation layer that screenshots and clicks, or a hand-built API wrapper that ends up covering maybe 10% of what the underlying tool can actually do. CLI-Anything skips both and bets on the command line itself — the argument being that text-in/text-out commands are already the format LLMs work in best, and that a --help flag plus structured JSON output gives an agent everything it needs to discover and use a tool. If that bet holds, you don't need vendors to ship "AI features" to make their products work with agents — you point a plugin at the codebase and get a working CLI on the other end.

The pragmatic question is how good those auto-generated CLIs really are. The project advertises a 7-phase pipeline — analyze the source, design command groups, implement, test, document, publish — and claims a 100% pass rate across 2,461 tests on dozens of generated harnesses. That's impressive on paper, but "passes its own tests" and "covers what an operator actually needs day-to-day" aren't the same thing. The /cli-anything:refine command exists precisely because the first pass leaves gaps you have to come back and fill. So the realistic mental model is less "one-shot conversion" and more "decent first draft you keep extending."

For a no-code operator, the directly useful angle isn't building your own CLIs from scratch — it's CLI-Hub, the registry of pre-built ones. Things like the n8n, Obsidian, Zotero, Exa, and AdGuard harnesses are already there, installable with one pip command, and immediately usable from Claude Code, OpenClaw, or any agent that can run shell commands. If you've ever wanted your agent to actually drive a tool instead of describing what it would do, this is the cheapest way to find out whether that works.

What it's good at

• Giving agents a real handle on desktop and self-hosted software — Blender, GIMP, OBS, LibreOffice, n8n, Obsidian, and dozens of others all have working CLI harnesses you can install today.
• Working with the agents you already use — first-class plugins for Claude Code and OpenClaw, plus community support for Codex, OpenCode, Goose, Pi, and GitHub Copilot CLI.
• Skipping the GUI-automation tax — no screenshots, no pixel-clicking, no flaky element selectors; it talks to the software's actual backend through commands.
• Shipping structured output by default — every generated CLI has a --json flag, so the agent gets clean structured data instead of having to parse human-readable tables.
• Discoverable without docs — the auto-generated SKILL.md plus --help means an agent can figure out what a CLI can do on its own, without you wiring up custom prompts.

What it's not

• Not a commercial product — academic project from the University of Hong Kong's Data Intelligence Lab (HKUDS); no support contract, no SLA, and the roadmap moves at the pace of the contributing community.
• Not a substitute for a focused, hand-built integration — if you need depth on one specific tool, a purpose-built integration will usually beat the auto-generated harness on coverage and reliability.
• Not always the right shape for visual tools — wrapping a creative tool like Blender in a CLI gets you scripted control, not the visual feedback loop most operators actually want. The provenance question — "is the CLI really representative of what the tool can do?" — is left to the agent and the user to figure out.
• Not zero-touch — generated CLIs need refinement passes to reach decent coverage, and you're on the hook for keeping them current as the underlying software changes.
• Not a replacement for the underlying app — you still need a working install of GIMP, n8n, Obsidian, etc.; CLI-Anything is a layer on top, not a reimplementation.