A Hong Kong AI research lab has released a tool that might quietly expand what coding agents can actually do more than any model update this month.
HKUDS—the group behind LightRAG and GraphRAG—published CLI-Anything this week, a Python framework that takes a software project as input and produces a fully tested, documented command-line wrapper. Point it at GIMP, Blender, LibreOffice, or most anything with a Python API, and it generates a CLI that agents can work with directly.
The problem is straightforward: coding agents like Claude Code, OpenAI Codex, and OpenCode are most effective through command-line interfaces, but most professional software ships with GUIs and patchy APIs. That leaves agents either wrestling with UI automation—brittle, slow, constantly breaking—or hitting walls where the API simply doesn't expose what they need.
CLI-Anything runs a seven-phase pipeline: analyze source code, design a Click-based CLI architecture, implement it with REPL mode, structured JSON output, and undo/redo support, then auto-generate tests and documentation before installing to PATH. Nothing ships until every test passes. The repository currently carries 1,508 passing tests across 11 demonstrated applications. A `refine` command lets agents deepen CLI coverage incrementally without touching existing functionality.
Distribution leans on Claude Code's plugin marketplace, giving the project immediate reach into what its README describes as a user base running 'thousands of real workflows through CLI daily.' Other listed platforms include OpenCode (from SST), Codex, Qodercli, nanobot, and OpenClaw—the agent layer inside AnyGen.io.
The tool's deeper bet is a strategic one. Writing bespoke MCP servers or browser automation scripts for individual applications doesn't scale—there's too much software and not enough engineers willing to maintain the glue. HKUDS is wagering that an automated, generalizable conversion layer is the more durable answer for the long tail of professional desktop and server software that vendors haven't adapted for agents. Whether generated CLIs hold up reliably across diverse, real-world codebases remains to be seen, but the test coverage suggests the lab is treating reliability as a first-class constraint rather than an afterthought.