CLI-Anything generates a command-line interface for an existing piece of software so that an AI agent can operate it without screenshots, UI automation, or a hand-written API wrapper. You point the generator at a local codebase, a GitHub repository, or a set of API docs, and it produces a Python package (cli-anything- ) whose commands call the real application for rendering and export: LibreOffice produces the PDF, Blender renders the scene, Audacity processes the audio.
The generator ships as a plugin or skill for coding agents, and the same seven-phase method runs in each: Claude Code (as a plugin marketplace), Cursor, Pi Coding Agent, OpenClaw, OpenCode, Codex, Hermes, Goose, Reasonix, Qodercli, and GitHub Copilot CLI. The phases cover codebase analysis, architecture design, implementation, test planning, test writing, documentation, and a SKILL.md for skill discovery, with refine, test, and validate commands for follow-up passes. A HARNESS.md playbook sets the rules every generated CLI follows: call the real backend, verify output by magic bytes and pixel or audio analysis rather than exit codes, and handle render-time effects and timecode rounding explicitly.
Every generated CLI runs in two modes: a stateful REPL with undo and redo for interactive agent sessions, and plain subcommands for scripts. A --json flag on each command returns structured output, and agents discover the tool through standard --help and which lookups. Completed CLIs are published to CLI-Hub, a registry and package manager where agents browse, install, and use community-built CLIs, with a meta-skill that points an agent at the live catalog.
Harness generation relies on frontier-class models such as Claude Opus 4.6, Claude Sonnet 4.6, or GPT-5.4. Quality depends on source code being available (compiled-only targets degrade coverage), and a first run often needs one or more refine passes. Covered software spans creative tools (GIMP, Blender, Inkscape, OBS Studio, Kdenlive), office and collaboration (LibreOffice, Zoom, Joplin), diagramming (Draw.io, Mermaid), AI platforms (ComfyUI, Ollama), game engines (Godot), testing (WireMock), GPU debugging (RenderDoc), fabrication (MeerK40t, Ink/Stitch), and web services such as Exa, n8n, and Dify.
Features
- CLI generation from source: analyses a local codebase, a GitHub repository, or API docs and produces a Python CLI package that drives the real software
- Seven-phase pipeline: architecture design, implementation, test planning, test writing, documentation, and SKILL.md generation run automatically, with refine, test, and validate commands afterward
- Real backend calls: generated CLIs write valid project files (ODF, MLT XML, SVG) and delegate rendering and export to the actual application, with no fallback when the backend is missing
- Dual interaction modes: a stateful REPL with undo and redo for agent sessions, plus subcommands for scripting
- Structured output: a --json flag on every command returns machine-readable data, and human-readable tables serve interactive use
- Many agent hosts: Claude Code, Cursor, Pi Coding Agent, OpenClaw, OpenCode, Codex, Hermes, Goose, Reasonix, Qodercli, and GitHub Copilot CLI
- CLI-Hub: a registry and package manager where agents discover, install, and manage community-built CLIs, with a meta-skill that points them at the live catalog
- HARNESS.md playbook: rules for authentic integration, filter translation between formats, timecode precision, and output verification by magic bytes and pixel or audio analysis
- Layered tests: unit tests, native end-to-end tests, true-backend tests against the real software, and subprocess tests of the installed command, with 2,464 passing across 30 harnesses
- API unification: wraps scattered web-service endpoints into one stateful CLI with grouped commands
