name: openrouter description: Invokes 300+ AI models via OpenRouter API for text completion, image generation, and model discovery. Use when delegating tasks to external models (GPT-5.2, Gemini 3, Llama, Mistral, etc.). Triggers on "use OpenRouter to...", "call GPT-5 to...", "generate an image with Gemini", or similar requests for external AI models. license: MIT compatibility: python 3.8+, requires requests library metadata: version: "1.1.1"

OpenRouter

Gateway to 300+ AI models through a unified API. Requires SKILL_OPENROUTER_API_KEY environment variable.

Setup

export SKILL_OPENROUTER_API_KEY="sk-or-..."  # Get key at https://openrouter.ai/keys

Sandbox Compatibility

⚠️ macOS Limitation: On macOS, uv run may require dangerouslyDisableSandbox: true because UV accesses system configuration APIs (SystemConfiguration.framework) to detect proxy settings. This is a known UV limitation on macOS systems.

Behavior:

• On first execution, Claude may attempt with sandbox enabled
• If it fails with system-configuration errors, Claude will retry with sandbox disabled
• This is expected behavior and does not indicate a security issue

Alternative (for restricted environments): If sandbox restrictions are problematic, you can pre-install dependencies:

python3 -m pip install requests
python3 /absolute/path/to/scripts/openrouter_client.py chat MODEL "prompt"

However, we recommend the standard UV approach for portability and zero-setup benefits.

Why UV is preferred:

• Zero setup required (no pip install step)
• Dependencies declared inline (PEP 723 standard)
• Automatic caching and fast execution
• Full portability across systems
• Official Anthropic/Claude Code recommendation

Quick Reference

Text completion:

UV_CACHE_DIR=/tmp/claude/uv-cache uv run --with requests scripts/openrouter_client.py chat MODEL "prompt" [--system "sys"] [--max-tokens N] [--temperature T]

Image generation:

UV_CACHE_DIR=/tmp/claude/uv-cache uv run --with requests scripts/openrouter_client.py image MODEL "description" [--output /absolute/path/file.png] [--aspect 16:9] [--size 2K] [--background transparent] [--quality high] [--output-format png]

Model discovery:

UV_CACHE_DIR=/tmp/claude/uv-cache uv run --with requests scripts/openrouter_client.py models [vision|image_gen|tools|long_context]
UV_CACHE_DIR=/tmp/claude/uv-cache uv run --with requests scripts/openrouter_client.py find "search term"

Common Models

Usage Patterns

Pattern 1: Sequential Model Chain

Call multiple models in sequence, passing outputs forward:

# Step 1: Generate outline with one model
outline = client.chat_simple("openai/gpt-5.2", "Create outline for: {topic}")

# Step 2: Expand with another model
content = client.chat_simple("anthropic/claude-sonnet-4.5", f"Expand this outline:\n{outline}")

Pattern 2: Parallel Model Comparison

Get responses from multiple models for comparison:

# Run these in parallel
UV_CACHE_DIR=/tmp/claude/uv-cache uv run --with requests scripts/openrouter_client.py chat openai/gpt-5.2 "Explain X" > gpt4_response.txt &
UV_CACHE_DIR=/tmp/claude/uv-cache uv run --with requests scripts/openrouter_client.py chat anthropic/claude-sonnet-4.5 "Explain X" > claude_response.txt &
wait

Pattern 3: Specialized Delegation

Route specific tasks to specialized models:

# Use code model for code tasks
UV_CACHE_DIR=/tmp/claude/uv-cache uv run --with requests scripts/openrouter_client.py chat anthropic/claude-sonnet-4.5 "Write a function to..."

# Use vision model for image analysis
UV_CACHE_DIR=/tmp/claude/uv-cache uv run --with requests scripts/openrouter_client.py chat google/gemini-3-flash-preview "Analyze this image: [base64]"

# Use image model for generation
UV_CACHE_DIR=/tmp/claude/uv-cache uv run --with requests scripts/openrouter_client.py image google/gemini-3-pro-image-preview "A cyberpunk city" -o city.png

Pattern 4: Structured Output Pipeline

Request JSON for programmatic processing:

# Get structured data
UV_CACHE_DIR=/tmp/claude/uv-cache uv run --with requests scripts/openrouter_client.py chat openai/gpt-5.2 "Extract entities from: {text}" --json > entities.json

# Process the JSON in next step

Image Generation

Required: Use chat completions endpoint with modalities: ["image", "text"]

Aspect ratios: 1:1, 16:9, 9:16, 4:3, 3:4, 21:9 Sizes: 1K, 2K, 4K

# Generate landscape image (use absolute path for --output)
UV_CACHE_DIR=/tmp/claude/uv-cache uv run --with requests scripts/openrouter_client.py image google/gemini-3-pro-image-preview \
  "Mountain sunset with dramatic clouds" \
  --output /absolute/path/to/mountain.png --aspect 16:9 --size 2K

Note: Always use absolute paths for --output to ensure images are saved to the correct location. The script creates parent directories automatically if they don't exist.

Advanced Image Generation Options

Transparent Backgrounds (GPT-5 Image and compatible models):

UV_CACHE_DIR=/tmp/claude/uv-cache uv run --with requests scripts/openrouter_client.py image \
  openai/gpt-5-image "Logo design with transparent background" \
  --background transparent \
  --quality high \
  --output-format png \
  --output /absolute/path/to/logo.png

New parameters:

• --background <value>: Background setting (e.g., transparent). Model support varies.
• --quality <value>: Image quality setting (high, medium, low). Affects detail and transparency quality.
• --output-format <format>: Output format (png, webp, jpeg). Use png or webp for transparency.

Transparency Requirements: For transparent backgrounds, specify transparency in BOTH:

1. API parameter: Use --background transparent flag
2. Prompt text: Include "transparent background" in the description

Both are required for best results. The --quality high and --output-format png parameters improve transparency quality.

Model Compatibility:

• ✅ openai/gpt-5-image - Full transparent background support
• ⚠️ Other models - Check model documentation for transparency support

Error Handling

The script handles retries automatically for transient errors (429, 502, 503).

Common errors:

• 401: Invalid API key - check SKILL_OPENROUTER_API_KEY
• 402: Add credits at openrouter.ai
• 429: Rate limited - script auto-retries

Python Usage (Direct Import)

For complex workflows, import the client directly:

import sys
sys.path.insert(0, "scripts")
from openrouter_client import OpenRouterClient
import os

client = OpenRouterClient(os.environ["SKILL_OPENROUTER_API_KEY"])

# Simple chat
response = client.chat_simple("anthropic/claude-sonnet-4.5", "Hello!")

# Full chat with history
result = client.chat("openai/gpt-5.2", [
    {"role": "system", "content": "You are a helpful assistant"},
    {"role": "user", "content": "Explain recursion"},
    {"role": "assistant", "content": "Recursion is..."},
    {"role": "user", "content": "Give an example"}
])
content = result["choices"][0]["message"]["content"]

# Generate image (use absolute path)
images = client.generate_image(
    "google/gemini-3-pro-image-preview",
    "A serene forest path",
    output_path="/absolute/path/to/forest.png",
    aspect_ratio="16:9"
)

# Generate image with transparent background
images = client.generate_image(
    "openai/gpt-5-image",
    "A minimalist logo with transparent background",
    output_path="/absolute/path/to/logo.png",
    aspect_ratio="1:1",
    background="transparent",
    quality="high",
    output_format="png"
)

# Find models
vision_models = client.list_models("vision")
claude_models = client.find_model("claude")