CLI-Anything Reference

Auto-generates production-ready, agent-controllable CLI interfaces for any GUI application with accessible source code. A 7-phase pipeline maps GUI actions to backend APIs and emits Click-based harnesses with --json output mode, interactive REPL, and pytest test suites.

Source: https://github.com/HKUDS/CLI-Anything License: MIT | Python: 3.10+ | Tests: 1,508 passing across 11 apps

---

Self-Evolving Skill: This skill improves through use. If instructions are wrong, parameters drifted, or a workaround was needed — fix this file immediately, don't defer. Only update for real, reproducible issues.

Installation

Method 1: Claude Code Marketplace (recommended)

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

Method 2: Manual Claude Code

git clone https://github.com/HKUDS/CLI-Anything.git
cp -r CLI-Anything/cli-anything-plugin ~/.claude/plugins/cli-anything
# Reload in Claude Code:
/reload-plugins

Method 3: OpenCode — global

git clone https://github.com/HKUDS/CLI-Anything.git
cp CLI-Anything/opencode-commands/*.md ~/.config/opencode/commands/
cp CLI-Anything/cli-anything-plugin/HARNESS.md ~/.config/opencode/commands/

Method 4: OpenCode — project-level

cp CLI-Anything/opencode-commands/*.md .opencode/commands/
cp CLI-Anything/cli-anything-plugin/HARNESS.md .opencode/commands/

---

Generating a New CLI Harness

# From local source path
/cli-anything ./gimp

# From GitHub URL (auto-cloned)
/cli-anything https://github.com/blender/blender

# Claude Code variant (plugin command)
/cli-anything ./inkscape

# OpenCode variant
/cli-anything ./libreoffice

---

All Commands

Command	Platform	Purpose
/cli-anything <path-or-url>	Claude Code	Build full 7-phase harness
/cli-anything:refine <path> [focus]	Claude Code	Expand with gap analysis
/cli-anything:test <path-or-url>	Claude Code	Run tests + update TEST.md
/cli-anything:validate <path-or-url>	Claude Code	Verify HARNESS.md standards
/cli-anything:list [--path dir] [--depth n] [--json]	Claude Code	Discover installed CLI-Anything tools
/cli-anything ./gimp	OpenCode	Build harness
/cli-anything-refine ./gimp	OpenCode	Refine harness
/cli-anything-test ./gimp	OpenCode	Run tests
/cli-anything-validate ./gimp	OpenCode	Validate standards
/cli-anything-list	OpenCode	List installed tools

Refinement with Focus

# Broad gap analysis
/cli-anything:refine ./gimp

# Targeted domain expansion
/cli-anything:refine ./gimp "batch processing and filters"
/cli-anything:refine ./blender "rendering pipeline and animations"
/cli-anything:refine ./libreoffice "spreadsheet formulas and macros"

---

Installing a Generated CLI

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

# Verify PATH entry created
which cli-anything-gimp
cli-anything-gimp --help

---

Using Generated CLIs

All generated CLIs share a consistent interface pattern.

GIMP Example

# Help
cli-anything-gimp --help

# Create project
cli-anything-gimp project new --width 1920 --height 1080 -o poster.json

# Machine-readable JSON output (for agent consumption)
cli-anything-gimp --json layer add -n "Background" --type solid --color "#1a1a2e"

# Interactive REPL
cli-anything-gimp

# Stateful workflow with --project flag
cli-anything-gimp --project poster.json layer add -n "Logo" --type group
cli-anything-gimp --project poster.json export render output.png --format png

LibreOffice Example

# New document
cli-anything-libreoffice document new -o report.json --type writer

# Populate document
cli-anything-libreoffice --project report.json writer add-heading -t "Q1 Report" --level 1
cli-anything-libreoffice --project report.json writer add-table --rows 4 --cols 3

# Export
cli-anything-libreoffice --project report.json export render output.pdf -p pdf --overwrite

# JSON output
cli-anything-libreoffice --json document info --project report.json

Blender Example

cli-anything-blender scene new -o scene.json
cli-anything-blender --project scene.json object add --type mesh --mesh-type CUBE
cli-anything-blender --project scene.json render image output.png --engine CYCLES
cli-anything-blender --json scene info --project scene.json

Inkscape Example

cli-anything-inkscape document new --width 800 --height 600 -o design.json
cli-anything-inkscape --project design.json shape add rect --x 10 --y 10 --w 200 --h 100
cli-anything-inkscape --project design.json export render output.svg
cli-anything-inkscape --json document info --project design.json

---

Output Modes

Every command supports two output modes:

Mode	Usage	Audience
Human-readable (default)	cli-anything-gimp layer add ...	Terminal / humans
Machine-readable JSON	cli-anything-gimp --json layer add ...	AI agents, scripts

The --json flag is always the first argument, before any subcommands.

---

Interactive REPL

Enter by running the CLI with no subcommands:

cli-anything-gimp
cli-anything-blender
cli-anything-libreoffice

REPL features:

• 50-level undo/redo stack
• Persistent command history
• ANSI 256-color output (software-specific accent colors)
• Table rendering, progress indicators
• All subcommands available interactively

---

Running Tests

# Standard pytest run
cd <software>/agent-harness
python3 -m pytest cli_anything/<software>/tests/ -v

# Validate the installed binary is invoked (not just the module)
CLI_ANYTHING_FORCE_INSTALLED=1 python3 -m pytest cli_anything/<software>/tests/ -v -s

# Coverage
python3 -m pytest cli_anything/<software>/tests/ --cov=cli_anything/<software>

# Per-software examples
cd gimp/agent-harness && python3 -m pytest cli_anything/gimp/tests/ -v
cd blender/agent-harness && python3 -m pytest cli_anything/blender/tests/ -v
cd libreoffice/agent-harness && python3 -m pytest cli_anything/libreoffice/tests/ -v

---

Environment Variables

Variable	Purpose
CLI_ANYTHING_FORCE_INSTALLED=1	Forces tests to use the real installed binary (validates which cli-anything-<software> resolves correctly, not just module import)

---

Supported Applications (11 total, 1,508 tests)

Application	Category	Entry Point
GIMP	Creative (raster)	cli-anything-gimp
Blender	Creative (3D)	cli-anything-blender
Inkscape	Creative (vector)	cli-anything-inkscape
Audacity	Media (audio)	cli-anything-audacity
LibreOffice	Productivity	cli-anything-libreoffice
OBS Studio	Media (streaming)	cli-anything-obs-studio
Kdenlive	Media (video)	cli-anything-kdenlive
Shotcut	Media (video)	cli-anything-shotcut
Zoom	Communication	cli-anything-zoom
Draw.io	Diagramming	cli-anything-drawio
AnyGen	AI Content	cli-anything-anygen

---

The 7-Phase Pipeline (HARNESS.md)

What /cli-anything triggers under the hood:

Phase	Name	What happens
1	Source Acquisition	Accept local path or GitHub URL; clone if needed
2	Codebase Analysis	Map GUI actions → backend APIs; inventory existing CLI tools; document data model
3	CLI Architecture Design	Choose interaction model (stateful REPL + subcommand); define command groups; plan --json output
4	Implementation	Build Click-based CLI with cli_anything.<software> namespace; implement REPL via shared ReplSkin
5	Test Planning	Write TEST.md with planned test inventory before code
6	Test Implementation	Unit tests (synthetic data) + E2E tests (real software output) + subprocess tests
7	PyPI Publishing	setup.py with find_namespace_packages(); console_scripts entry points

Fundamental Rule: The CLI MUST delegate rendering to the real software — never reimplement functionality in Python. The target software is a hard dependency.

---

Generated Harness Structure

Each generated CLI follows this layout (PEP 420 namespace packages):

<software>/agent-harness/
├── setup.py                          # find_namespace_packages(), entry_points
├── <SOFTWARE>.md                     # Software-specific SOP
└── cli_anything/                     # NO __init__.py (namespace package)
    └── <software>/
        ├── __init__.py
        ├── __main__.py
        ├── <software>_cli.py         # Main Click entry point
        ├── core/                     # Domain modules (project, session, export, ...)
        ├── utils/
        │   └── repl_skin.py          # Shared REPL UI (copied from plugin)
        └── tests/
            ├── TEST.md               # Test inventory (written in phase 5)
            ├── test_core.py          # Unit tests
            └── test_full_e2e.py      # E2E + subprocess tests

Why namespace packages: cli_anything/ has no __init__.py, enabling multiple separately-installed CLIs to coexist under the same cli_anything.* namespace without conflict.

---

Output Verification Standards

Tests verify real output using magic bytes and content analysis:

Output Type	Verification
PDF	%PDF- magic bytes + non-zero size
OOXML (.docx, .xlsx)	ZIP structure validation
CSV	Data row count + column check
Images	Pillow pixel value analysis
Audio	RMS level check (non-silent)

---

Prerequisites Checklist

Before running /cli-anything <path>:

• Python 3.10+ installed
• Target software installed (hard dependency — it will be called for rendering)
• Source code accessible (local path or public GitHub URL)
• click>=8.0.0 available (installed via pip install -e .)
• Pillow>=10.0.0, prompt-toolkit>=3.0.0 available
• pytest available for running tests

---

Key Files Reference

File	Purpose
cli-anything-plugin/HARNESS.md	Complete methodology document (30K chars)
cli-anything-plugin/QUICKSTART.md	Quick start guide
cli-anything-plugin/repl_skin.py	Shared REPL UI class (18K, ANSI 256-color)
cli-anything-plugin/commands/cli-anything.md	/cli-anything slash command definition
cli-anything-plugin/commands/refine.md	/cli-anything:refine command
cli-anything-plugin/commands/test.md	/cli-anything:test command
cli-anything-plugin/commands/validate.md	/cli-anything:validate command
cli-anything-plugin/commands/list.md	/cli-anything:list command

Post-Execution Reflection

After this skill completes, check before closing:

1. Did the command succeed? — If not, fix the instruction or error table that caused the failure.
2. Did parameters or output change? — If the underlying tool's interface drifted, update Usage examples and Parameters table to match.
3. Was a workaround needed? — If you had to improvise (different flags, extra steps), update this SKILL.md so the next invocation doesn't need the same workaround.

Only update if the issue is real and reproducible — not speculative.