HKUDS/CLI-Anything

A generator that wraps existing software in structured, agent-friendly command-line interfaces so AI coding agents can drive it.

What it is

CLI-Anything is a Python toolchain for turning arbitrary software — desktop apps, backend services, or source repositories — into command-line harnesses that AI agents can call. It targets developers who want their AI coding agent (Claude Code, Codex, OpenClaw, and others) to operate tools that only expose a GUI or an API by giving those tools a consistent, self-describing CLI. Alongside the generator it ships CLI-Hub, a package manager for browsing, installing, and managing community-built CLIs.

Key features

A seven-phase generator pipeline (analyze, design, implement, plan tests, write tests, document, publish) that produces a Click-based Python CLI with a REPL, JSON output, and undo/redo.
CLI-Hub package manager (pip install cli-anything-hub) with list, search, info, install, update, uninstall, and launch commands for both first-party and public third-party CLIs.
Every generated CLI ships a SKILL.md definition so skills-compatible agents can discover and install it.
Plugin or skill integrations for Claude Code, Pi, OpenCode, OpenClaw, Codex, Hermes, Reasonix, Qodercli, GitHub Copilot CLI, and Goose.
A refine command that performs gap analysis against a target's full capabilities and incrementally adds commands, tests, and docs.

Tech stack

Python 3.10+ (primary language).
Click 8.0+ for the generated CLIs.
pytest for the unit and end-to-end test suites; generated harnesses publish via setup.py.
Distributed as a Claude Code plugin marketplace (.claude-plugin/marketplace.json) and as per-agent skill and extension installers.

When to reach for it

You want an AI agent to operate desktop software (GIMP, Blender, LibreOffice, QGIS, FreeCAD, and similar) that has no agent-native interface.
You need a repeatable way to generate and test CLIs for internal tools or codebases.
You already work inside a supported agent platform and want a one-command way to add tool coverage.
You want to install existing community CLIs rather than build your own.

When not to reach for it

Your target software already exposes a stable API or MCP server that agents can call directly.
You are not working through an AI coding agent — the value is agent-driven CLI generation, not hand-authored tools.
You need a language other than Python for the generated harness; output is Click and Python.
You want a single small dependency; this is a large, actively growing monorepo of many harnesses.

Maturity signal

Created in early March 2026 and pushed within the last few weeks, the project is very young but moving quickly, with a large star count and a steady stream of merged CLI harnesses and fixes documented in the README's news log. Apache-2.0 licensing and the volume of contributor-submitted harnesses point to an actively maintained, community-driven effort rather than a settled, stable release. Expect rapid change and evolving conventions.

Alternatives

Model Context Protocol servers — use MCP directly when your tool can expose a standardized protocol endpoint instead of a generated CLI.
Open Interpreter — use it when you want an agent to write and run code against a tool ad hoc rather than through a pre-built, tested harness.
Hand-written Click or Typer CLIs — use these when you need full control over a single tool and do not want a generator's conventions.

Notes

The README documents a large and growing catalog of harnesses (GIMP, Blender, QGIS, FreeCAD with 258 commands, Calibre, Zotero, and many more) plus repeated security hardening passes (path-traversal fixes, defusedxml for untrusted XML), suggesting the generated CLIs are treated as real software with test and security expectations rather than throwaway wrappers.

Wiki: HKUDS/CLI-Anything