Cli Anything

cli-anything — Make Any Software Agent-Native

CLI-Anything (HKUDS) bridges AI agents and the world's software: instead of brittle GUI automation, it wraps real applications (Blender, GIMP, LibreOffice, OBS, Godot, QGIS, …) in structured, agent-first CLIs — Click commands, stateful REPL, --json output, undo/redo — generated and validated through a 7-phase pipeline with 2,461 passing tests across 40+ harnesses. This skill is the routing-first wrapper: pick the right mode, install the right surface, and drive the build → refine → validate loop.

When to use this skill

• The user wants an agent to control real software (image/video/audio

editing, office, 3D, GIS, game engines, diagramming, streaming) through a CLI instead of screenshots or RPA

• The user asks to install cli-anything-hub / use cli-hub

(list, search, info, install, update, uninstall, launch)

• The user wants agents to autonomously discover and install CLIs via

the CLI-Hub meta-skill

• The user wants to generate a new CLI harness from a local codebase or

GitHub repo with /cli-anything, or asked how to install the CLI-Anything plugin/skill for Claude Code, Codex, OpenCode, OpenClaw, Pi, Hermes, Qodercli, or GitHub Copilot CLI

• The user wants to expand or verify an existing harness with

/cli-anything:refine, /cli-anything:test, or /cli-anything:validate

When not to use this skill

• The user wants to design a domain-specific agent team (agents +

skills scaffolding), not a software CLI wrapper → use harness

• The user is browsing/installing Claude plugins generally → use

ccpi-marketplace

• The target has no codebase and no API and must be driven through a

real browser/GUI session → use browser-harness (or playwriter for an already-open browser)

• The user only needs web scraping/extraction → use scrapling

Prerequisites

Instructions

Step 1 — Choose the mode

Prefer the registry first — building is the fallback, not the default.

Step 2 — Use the ecosystem via CLI-Hub

pip install cli-anything-hub

cli-hub list                 # browse the registry
cli-hub search image         # search by keyword
cli-hub info gimp            # inspect one CLI
cli-hub install gimp         # install from CLI-Hub
cli-hub launch gimp          # run an installed CLI
cli-hub update gimp          # update
cli-hub uninstall gimp       # remove

CLI-Hub v0.2.0+ also serves public CLIs from multiple install sources (pip, npm, brew, bundled/system tools) via public_registry.json. If a harness wraps desktop software (GIMP, Blender, LibreOffice, …), install that upstream application too.

Step 3 — Give agents the meta-skill

npx skills add HKUDS/CLI-Anything --skill cli-hub-meta-skill -g -y

Then prompt the agent:

Find appropriate CLI software in CLI-Hub and complete the task: ...

Works with Claude Code, Codex, OpenClaw, Nanobot, Antigravity, and other SKILL-compatible agents. The monorepo's per-harness skills are also discoverable: npx skills add HKUDS/CLI-Anything --list.

Step 4 — Generate a new harness (7-phase pipeline)

Claude Code (primary platform):

/plugin marketplace add HKUDS/CLI-Anything
/plugin install cli-anything

/cli-anything ./gimp                              # local source
/cli-anything https://github.com/blender/blender  # GitHub repo

The pipeline runs: 1 Analyze → 2 Design → 3 Implement (Click CLI + REPL + JSON + undo/redo) → 4 Plan tests → 5 Write tests → 6 Document (+ Phase 6.5 SKILL.md generation) → 7 Publish (setup.py, PATH install). Other platforms (Codex, OpenCode, OpenClaw, Pi, Hermes, Qodercli, Copilot CLI, Goose) follow the same methodology — install commands are in `references/commands.md`.

If /cli-anything is unrecognized: /reload-plugins, verify with /help cli-anything, reinstall from the marketplace, and only on older builds try the legacy /cli-anything:cli-anything form.

Step 5 — Refine, test, validate

/cli-anything:refine ./gimp                                  # broad gap analysis
/cli-anything:refine ./gimp "batch processing and filters"   # focused
/cli-anything:test ./gimp                                    # run tests, update TEST.md
/cli-anything:validate ./gimp                                # check against HARNESS.md

Refinement is incremental and non-destructive — run it repeatedly to push coverage toward production quality. HARNESS.md is the methodology SOP (source of truth) the validator checks against.

Step 6 — Use a generated CLI

cd <software>/agent-harness && pip install -e .

cli-anything-gimp --help                     # discoverable via --help/which
cli-anything-gimp --json layer add -n "BG"   # structured output for agents
cli-anything-gimp                            # bare command → stateful REPL

Every harness ships dual modes (REPL + subcommands), --json on every command, and a generated SKILL.md at skills/cli-anything-<name>/SKILL.md.

Step 7 — Plugin-style installation alongside jeo-skills

This skill folder is plugin-installable through the standard jeo-skills flow so the wrapper, references, and installer script land on disk for any supported agent runtime:

# Project install (writes into .agents/skills/cli-anything/)
npx skills add https://github.com/akillness/jeo-skills --skill cli-anything

# Global install for every detected agent
npx skills add -g https://github.com/akillness/jeo-skills --skill cli-anything

# Target specific agents
npx skills add -g https://github.com/akillness/jeo-skills --skill cli-anything -a claude-code -a codex -y

The skill also ships `scripts/install.sh` — a one-shot CLI-Hub installer (venv-aware uv/pip selection, PEP 668-safe) with an optional meta-skill install via CLI_ANYTHING_META_SKILL=1.

Output format

When the user asks cli-anything for help, return a compact brief:

# cli-anything Routing Brief

## Scope
- Mode: use-ecosystem | empower-agents | generate-harness | iterate-harness
- Target software: <name or repo> | undecided
- Agent platform: claude | codex | opencode | openclaw | pi | hermes | copilot | other

## Recommended next move
- install-cli-hub | cli-hub-search | install-meta-skill | install-platform-plugin | run-/cli-anything | refine | test | validate

## Why
- 2-3 bullets grounded in the user's packet

## Route-outs
- `harness` for agent-team architecture (not software wrappers)
- `browser-harness` for no-codebase GUI/browser targets
- `ccpi-marketplace` for general Claude plugin browsing

Best practices

1. Search the registry before generating — 40+ harnesses already exist; cli-hub search costs seconds, a 7-phase build costs a session. 2. Install the real backend — harnesses call actual software with zero compromise dependencies; tests fail (not skip) when backends are missing. 3. Use `--json` for agent consumption — every command supports it; human-readable tables are for debugging only. 4. Verify outputs, not exit codes — the methodology checks magic bytes, ZIP/OOXML structure, pixel/audio analysis; do the same with artifacts. 5. Run `/cli-anything:refine` iteratively — a single generation pass rarely covers all capabilities; each refine run is incremental. 6. Windows needs `bash` — install Git for Windows (bash + cygpath) or use WSL before running plugin commands.

References

• Upstream repo: <https://github.com/HKUDS/CLI-Anything>
• CLI-Hub (web): <https://hkuds.github.io/CLI-Anything/>
• Methodology SOP: `cli-anything-plugin/HARNESS.md`
• Quickstart: `cli-anything-plugin/QUICKSTART.md`
• Publishing guide: `cli-anything-plugin/PUBLISHING.md`
• Technical report: <https://arxiv.org/abs/2606.03854>
• Installer script: `scripts/install.sh`
• Command + platform reference: `references/commands.md`
• Adjacent skills: ../harness/SKILL.md, ../ccpi-marketplace/SKILL.md,

../browser-harness/SKILL.md, ../scrapling/SKILL.md

• License: Apache-2.0 (see upstream LICENSE)

cli-anything command and platform reference

Upstream: <https://github.com/HKUDS/CLI-Anything> · Hub: <https://hkuds.github.io/CLI-Anything/>

CLI-Hub package manager (cli-hub)

v0.2.0+ supports public CLIs from multiple install sources (pip, npm, brew, bundled/system tools) via public_registry.json.

Agent meta-skill (autonomous discovery)

npx skills add HKUDS/CLI-Anything --skill cli-hub-meta-skill -g -y
# Per-harness skills are also listable:
npx skills add HKUDS/CLI-Anything --list

Also published on ClawHub, SkillHub, and SkillHub.cn.

Generator commands (run inside the agent platform)

7-phase pipeline: Analyze → Design → Implement (Click + REPL + JSON + undo/redo) → Plan Tests → Write Tests → Document (+ 6.5 SKILL.md generation via skill_generator.py) → Publish (setup.py, PATH).

Platform install matrix

Claude Code troubleshooting (Unknown skill: cli-anything): /reload-plugins → /help cli-anything → reinstall from marketplace → legacy form /cli-anything:cli-anything only on older builds. Windows: requires Git for Windows (bash + cygpath) or WSL.

Generated harness usage

cd <software>/agent-harness && pip install -e .   # installs cli-anything-<name> on PATH

cli-anything-<name> --help        # self-describing
cli-anything-<name> --json <cmd>  # structured output for agents
cli-anything-<name>               # bare command enters stateful REPL

# Tests (force-installed mode recommended for validation)
CLI_ANYTHING_FORCE_INSTALLED=1 python3 -m pytest cli_anything/<name>/tests/ -v -s

Harness catalog (selected, 40+ total — 2,461 tests, 100% pass)

Creative/media: GIMP (107) · Blender (208) · Inkscape (202) · Audacity (161) · Krita · Kdenlive (155) · Shotcut (154) · Openscreen (101) · OBS Studio (153) · MuseScore (56) · VideoCaptioner (26) · Sketch (19) Office/knowledge: LibreOffice (158) · Zotero · Calibre (58) · Joplin (134) · Mubu (96) · Obsidian · NotebookLM (21) Diagrams/GIS/CAD: Draw.io (138) · Mermaid (10) · QGIS (22) · FreeCAD · 3MF (50) · CloudCompare (88) · CloudAnalyzer (14) AI/dev/infra: ComfyUI (70) · Ollama (98) · Exa (40) · n8n · Dify Workflow (11) · Mailchimp (303 cmds) · AdGuard Home (36) · WireMock · ETH2 QuickStart (18) · LLDB (27) · RenderDoc (59) · Nsight Graphics (51) · Unreal Insights (50) Games: Godot (24) · s&box (244) · Slay the Spire II Comms: Zoom (22) · Browser (DOMShell MCP)

Design principles (from HARNESS.md)

1. Authentic integration — generate valid project files (ODF, MLT XML, SVG), delegate rendering to the real application; interfaces TO software, not replacements. 2. Dual interaction — stateful REPL for agent sessions + subcommands for pipelines; bare command enters REPL. 3. Agent-native output — --json on every command; discovery via --help and which. 4. Zero-compromise backends — real software is a hard requirement; tests fail (not skip) when backends are missing. 5. Verify artifacts — magic bytes, ZIP/OOXML structure, pixel/audio analysis, duration checks — never trust exit code 0.

Limitations

• Harness generation needs frontier-class models; weaker models produce

incomplete CLIs.

• Works from source code; binary-only targets degrade substantially.
• One generation pass rarely covers everything — plan for /cli-anything:refine.