HKUDS/CLI-Anything

The problem

Agents are good at shell tools and bad at everything else. Most real software -- CAD, video editors, DAWs, GIS, note apps -- exposes no stable programmatic interface, so an agent that can write code still cannot open Blender and produce a render. CLI-Anything's answer is a registry of per-application CLI harnesses plus a hub (cli-hub install <name>) that puts a scriptable, JSON-friendly surface in front of each app, so the agent calls commands instead of clicking.

What's different

MCP servers are the usual way to give an agent access to an app, and they work, but each server is bespoke, often stateful, and hard to test from a terminal. CLI-Anything standardizes on the CLI as the agent interface: subcommands, JSON plus human output, and a packaged SKILL.md describing usage. That is a narrower, more inspectable contract than a general-purpose MCP tool surface, and it composes with the existing skills ecosystems (npx skills add ...). The trade-off is that every app still needs someone to write and maintain a harness, and behaviour depends on how well the underlying app can be scripted (headless modes, local REST APIs, vendor CLIs).

Who it's for

Teams running coding agents (Claude Code, Cursor and similar) who want those agents to touch creative and productivity tools, not just repositories. Also useful for people who prefer driving an app from a shell with reproducible, testable commands. Look elsewhere if you need a single deeply integrated plugin for one app, or if your workflow is pure text/code -- the harness layer adds nothing there.

Cost of adoption

Python >=3.10, Apache-2.0, installed via pip for the hub. Per-harness requirements vary a lot: some wrap a local app on macOS or Windows, some need vendor credentials or tokens, some need the target app installed. Maturity is uneven by harness -- the README's news feed shows a constant stream of merges, security hardening (token handling, defusedxml for untrusted input) and E2E test additions, which is a good sign but also means quality differs between entries. Expect to read the individual harness docs and verify against your app version.

Verdict

This is a young project by age (created 2026-03-08) but not by attention: roughly 56.8k stars and +3,303 in a single week, ranked #14 on the weekly board. That pattern says a large community formed fast around the registry model, not that every harness is polished. The idea -- CLI as the agent-native interface -- is practical and low-magic, and the test/security discipline visible in recent PRs is more than most hype projects show. Worth adopting for a specific app whose harness is well tested; not worth adopting wholesale as infrastructure without checking the harness you need.