api-anything 0.2.0

Auto-generate Python APIs from any software. Beyond CLI-Anything.

API-Anything

Auto-generate Python APIs from any software. Beyond CLI.

CLI-Anything generates CLIs. We generate Python APIs -- typed, tested, deterministic.

---

Why API-Anything?

• For AI agents: A Python API call costs ~50 tokens. A CLI invocation costs ~200. Type hints eliminate hallucinated flags.
• For developers: import ffmpeg_api with autocomplete beats memorizing CLI flags and parsing stderr.
• For unreachable software: UIA mode controls software with NO source code and NO CLI. CLI-Anything cannot do this.

---

CLI-Anything vs API-Anything

Feature	CLI-Anything	API-Anything
Output format	CLI commands	Python API classes
Token cost per call	~200 tokens	~50 tokens
Type safety	None	Full type hints + Pydantic
IDE autocomplete	No	Yes
State management	Stateless (new process each call)	Stateful (class instance)
Error handling	Exit codes + stderr text	Python exceptions with context
Information sources	Source code only	Source code + UIA + screenshots
No-source-code software	Limited	Full UIA support (Tier 3)
Tier detection	No	7-level system saves tokens
Runtime AI needed	No	No (deterministic generated code)

Both tools solve the same problem from different angles. CLI-Anything (31k stars) wraps software as CLI commands. API-Anything wraps software as importable Python modules. They are complementary -- CLI-Anything covers breadth, API-Anything covers depth and type safety.

---

Tier Detection System

API-Anything classifies any software into 7 tiers before generating code, so it picks the right strategy and estimates cost upfront.

Tier	Rating	Strategy	Confidence
1	Stars: 5/5	Python binding exists on PyPI -- wrap directly	98%
2A	Stars: 4/5	Source code + CLI available -- subprocess mode	90%
2B	Stars: 3/5	Source code available, no CLI	75%
3A	Stars: 3/5	No source, UIA >=70% named controls	70%
3B	Stars: 2/5	No source, UIA 30-70% named controls	50%
3C	Stars: 1/5	No source, UIA <30% named controls	30%
4	Stars: 1/5	Canvas/game/custom-rendered -- vision required	15%

$ api-anything detect notepad

+------------------------------ api-anything ------------------------------+
|                                                                          |
|  [*] API-Anything Tier Detection                                         |
|     Software: notepad                                                    |
|     Tier: 1 *****                                                        |
|     Confidence: 98%                                                      |
|     Verdict: Direct binding -- wrap and ship!                            |
|                                                                          |
+--------------------------------------------------------------------------+
+----------------------------- Detection Checks ----------------------------+
|  [+] PyPI binding     python-notepad v0.1.2                              |
|  [--] Source code      no source path or URL provided                    |
|  [+] CLI binary       C:\Windows\System32\notepad.exe                    |
|  [--] UIA tree         uiautomation not installed -- skipped             |
+--------------------------------------------------------------------------+

---

Quick Start

pip install api-anything            # core
pip install "api-anything[all]"     # + web + mcp + test + repl

# First-run interactive wizard (recommended)
api-anything init

# List 27 built-in APIs (FFmpeg, Blender, Ollama, Docker, Git, SQLite, ...)
api-anything list --detail

# Detect tier for any software
api-anything detect notepad
api-anything detect gimp --github https://github.com/GIMP/gimp

# Branded stateful REPL (banner + state prompt + accent color per API)
api-anything repl examples/ffmpeg_api.py
# ffmpeg[input.mp4]> trim(0.0, 10.0)

# One-shot execution (piping friendly)
api-anything run SQLite get_tables
echo '{"x": 42}' | api-anything run jq query ".x"

# MCP server -- expose all 27 APIs to any MCP client (Claude Desktop, etc.)
api-anything mcp-serve --transport stdio

# Shell completion (bash/zsh/fish)
eval "$(api-anything completions bash)"

# Iteratively expand, test, validate a generated API
api-anything refine path/to/api.py --max-iter 3
api-anything test path/to/api.py
api-anything validate path/to/api.py

# Gradio demo web UI (onboarding / sales pitch)
api-anything web --port 7860

# Direct Python import
python -c "from api_anything.apis import SQLite; print(SQLite(':memory:').get_tables())"

Claude Code plugin

/plugin install api-anything

Provides /api-anything:detect, /api-anything:skill, /api-anything:refine, /api-anything:test, /api-anything:validate, /api-anything:list, /api-anything:repl, /api-anything:generate slash commands out of the box.

---

POC Showcase

LibreOffice Writer (Tier 2A -- subprocess mode)

from examples.libreoffice_writer import LibreOfficeWriter

writer = LibreOfficeWriter()
writer.create_document("report.odt")
writer.add_heading("Q1 Report", level=1)
writer.add_paragraph("Revenue grew 15% year-over-year.")
writer.add_table(
    headers=["Month", "Revenue"],
    rows=[["Jan", "$1M"], ["Feb", "$1.2M"]],
)
writer.save()
writer.export_pdf("report.pdf")

FFmpeg (Tier 2A -- subprocess mode)

FFmpeg has no Python binding on PyPI that wraps the full CLI. API-Anything generates one.

from examples.ffmpeg_api import FFmpeg

ff = FFmpeg()
ff.convert("input.mp4", "output.webm")
ff.extract_audio("video.mp4", "audio.mp3")
ff.trim("input.mp4", "clip.mp4", start="00:01:00", duration="00:00:30")

info = ff.probe("input.mp4")
print(f"Duration: {info['duration']}s, Resolution: {info['width']}x{info['height']}")

Windows Notepad (Tier 3A -- UIA mode)

This is what CLI-Anything cannot do. No source code, no CLI flags -- pure UI Automation.

from examples.notepad_uia_api import NotepadAPI

notepad = NotepadAPI()
notepad.launch()
notepad.set_text("Hello from API-Anything!")
notepad.save_as("hello.txt")
text = notepad.get_text()
notepad.close()

---

How It Works

API-Anything runs an 8-phase pipeline to turn any software into a Python package:

Phase 0: Tier Detection
   |  Classify software into 7 tiers (PyPI / source / UIA / vision)
   v
Phase 1: Three-Source Analysis
   |  Extract info from source code + UIA tree + screenshots
   v
Phase 2: API Surface Design
   |  LLM designs class/method structure with Instructor + Pydantic
   v
Phase 3: Code Generation
   |  Generate typed Python module with docstrings
   v
Phase 4: Validation
   |  Type-check + lint + smoke test generated code
   v
Phase 5: Test Generation
   |  Generate pytest suite for the API
   v
Phase 6: Packaging
   |  Create installable pip package
   v
Phase 7: Documentation
      Generate usage examples and API reference

Phases 0-1 are implemented. Phases 2-3 have core infrastructure (generator engine + Instructor integration). Phases 4-7 are on the roadmap.

---

Three-Source Analysis

API-Anything gathers information from three independent sources to maximize coverage:

1. Source code -- Parse repos or local code for function signatures, docstrings, and architecture. Works for open-source software (Tier 2A/2B).

2. UIA (UI Automation) -- Walk the Windows accessibility tree to discover controls, their types, names, and automation IDs. Works for closed-source GUI software (Tier 3A/3B/3C).

3. Screenshots -- Capture the UI for vision-model analysis when UIA coverage is low. Fallback for Tier 3C and Tier 4.

Each source fills gaps the others miss. Source code gives intent; UIA gives structure; screenshots give layout.

---

Architecture

API-Anything (open source, Apache-2.0)
  |
  +-- Tier Detector       classify software feasibility
  +-- Three-Source Analyzer   UIA + source + screenshots
  +-- Generator Engine     Instructor + Pydantic structured output
  +-- Generated Package    pure Python, no AI at runtime
       |
       +-- Typed API class    autocomplete + type hints
       +-- Python exceptions  not exit codes
       +-- pytest suite       generated tests
       +-- pip installable    any agent can import

The generated code is deterministic. No LLM runs at call time -- AI is used only during generation.

---

Roadmap

• Phase 0 -- Tier Detection (7-level system)
• Core generator engine (Instructor + Pydantic)
• POC: LibreOffice Writer (Tier 2A, subprocess)
• POC: FFmpeg (Tier 2A, subprocess)
• POC: Windows Notepad (Tier 3A, UIA)
• Three-source analyzer (UIA + source + screenshots)
• Gradio web demo (4 tabs: Dashboard / Tier / Explorer / Generate)
• SKILL.md generator (api-anything skill <module>)
• 27 built-in APIs (295 methods) -- FFmpeg, Blender, GIMP, Ollama, etc.
• Branded stateful REPL (banner + state prompt + accent color + 50-level undo/redo)
• Claude Code plugin (/plugin install api-anything)
• MCP server (api-anything mcp-serve)
• Shell completion generator (bash/zsh/fish)
• Interactive setup wizard (api-anything init)
• Refine / test / validate iterative pipeline
• stdin/stdout pipe support
• tool_use schema output (for any LLM agent)
• PyPI package publish (awaiting release cut)
• AT-SPI (Linux) and AXUIElement (macOS) UIA backends

---

Contributing

Contributions are welcome. The project is in early alpha (v0.1.0).

git clone https://github.com/anthropics/api-anything.git
cd api-anything
pip install -e ".[dev]"
pytest

Areas where help is needed:

• New POCs for popular software (Blender, Photoshop, Excel)
• UIA coverage improvements on Linux (AT-SPI) and macOS (AXUIElement)
• Phase 4-7 pipeline implementation

---

License

Apache-2.0. See LICENSE for details.

---

Credits

CLI-Anything (HKUDS, 31k stars) pioneered the idea of auto-generating agent-native interfaces for arbitrary software. API-Anything takes a complementary approach: generating importable, typed Python APIs instead of Click CLIs, with three-source analysis (UIA + screenshot + source code) so we reach the Tier 3 closed-source GUI software CLI-Anything cannot.

---

Generated with API-Anything v0.1.0