| name | model-scout |
| description | Discovers and recommends AI models via the OpenRouter API. Matches task requirements (cost, speed, capabilities, context length, modality) to available models. Use when the user needs to choose a model, compare models, validate model IDs, or when selecting a model for a new script. |
Model Scout
Discover and recommend AI models from OpenRouter based on task requirements.
Activation
This skill activates when user:
- Asks "which model should I use for X"
- Wants to compare models side-by-side
- Needs the cheapest/fastest/best model for a task
- Wants to validate that model IDs in
ai-models.md are still current
- Is building a new script and needs to select a model
Operations
| Operation | Description | Command |
|---|
| Recommend | Score and rank models for a task | recommend |
| Compare | Side-by-side comparison of specific models | compare |
| Audit | Validate and improve ai-models.md | audit |
Quick Start
uv run .claude/skills/model-scout/scripts/scout.py recommend \
--task "summarize 500 documents cheaply"
uv run .claude/skills/model-scout/scripts/scout.py recommend \
--task "agent with tool calling" \
--needs tools,reasoning \
--min-context 128000 \
--top 5
uv run .claude/skills/model-scout/scripts/scout.py compare \
--models anthropic/claude-sonnet-4 google/gemini-2.5-flash-preview openai/gpt-4o
uv run .claude/skills/model-scout/scripts/scout.py audit --suggest-updates
CLI Arguments
recommend
| Argument | Type | Default | Description |
|---|
--task | str | required | Natural language task description |
--needs | str | none | Required capabilities (comma-separated): tools, reasoning, vision, structured_outputs, audio |
--min-context | int | 0 | Minimum context window in tokens |
--max-cost | float | none | Max combined cost per 1M tokens (prompt + completion USD) |
--top | int | 5 | Number of results to show |
--all-providers | flag | false | Remove core-three provider bias for full exploration |
--json | flag | false | Output as JSON |
compare
| Argument | Type | Default | Description |
|---|
--models | str[] | required | 2-5 model IDs to compare |
--json | flag | false | Output as JSON |
audit
| Argument | Type | Default | Description |
|---|
--suggest-updates | flag | false | Print a ready-to-paste replacement table |
--json | flag | false | Output as JSON |
How Scoring Works
Models are scored across 5 dimensions (0-100 each) with dynamic weights:
| Dimension | What | Default Weight |
|---|
| Provider tier | Core three (OpenAI/Google/Anthropic) = 100, others = 50 | 15% |
| Cost | Log-scale inverse of prompt+completion per 1M tokens | 25% |
| Context | Size relative to requirement | 15% |
| Capabilities | Required params met + bonus for extras | 20% |
| Quality | Name-based tier lookup (opus=95, sonnet=80, flash=65, etc.) | 25% |
Weights shift based on task signal words (e.g., "cheap" boosts cost weight to 35%).
Environment
| Variable | Description |
|---|
OPENROUTER_API_KEY | API key from OpenRouter |
