Skilled Models Advisor

Skilled Models Advisor

Provides access to the tool-models agent and its query CLI for data-driven model selection. All recommendations default to models we have active credentials for. Pass --all to any command to see the full 2,700+ model universe.

---

When to use this skill

Consult tool-models before:

• Spawning a sub-agent — pick the right model for the task at hand
• User asks "which model for X" — always query rather than guess
• Estimating costs — especially before high-volume or recurring workloads
• Recommending a config change — primary/fallback assignments
• Comparing providers for the same model — context limits and output caps differ
• Checking if a model supports vision, PDF, reasoning, or function calling

Do NOT guess model specs from training data. Prices change, new models appear, and provider limits (output caps, context) vary from the model's stated maximum.

---

Spawn syntax

// Minimal — natural language question
sessions_spawn({
  agentId: "tool-models",
  runtime: "subagent",
  task: "Which model is best for processing long legal PDFs on a $50/month budget?"
})

// With structured context
sessions_spawn({
  agentId: "tool-models",
  runtime: "subagent",
  task: "Compare claude-sonnet-4-6 vs gemini-2.5-flash for a high-volume summarization pipeline. We process ~2M input tokens and ~200k output tokens per day."
})

// Model selection for a spawn
sessions_spawn({
  agentId: "tool-models",
  runtime: "subagent",
  task: "I'm about to spawn a coding sub-agent for refactoring a large CFML codebase. What's the best accessible model balancing code quality and cost?"
})

---

Direct CLI (when running in tool-models workspace)

QUERY="python3 $HOME/.openclaw/workspace-tool-models/skills/model-advisor/scripts/query.py"

# Stats — what's in the database, what we have access to
$QUERY stats

# Task-based recommendation (accessible models only by default)
$QUERY recommend "summarize long PDFs cheaply"
$QUERY recommend "complex reasoning and code generation"
$QUERY recommend "real-time low-latency chat"
$QUERY recommend "process images and extract structured data"

# Structured filters
$QUERY filter --reasoning --available                       # reasoning models in OpenClaw
$QUERY filter --cost-max 0.5 --ctx-min 128000              # under $0.50/M, 128k+ context
$QUERY filter --vision --pdf --cost-max 1                  # vision+PDF under $1/M
$QUERY filter --cost-max 0                                 # free models only

# Side-by-side comparison
$QUERY compare gpt-4.1 claude-sonnet-4-6 gemini-2.5-flash
$QUERY compare mistral-medium-latest mistral-small-latest

# Cost estimation
$QUERY cost gemini-2.5-flash --input 1000000 --output 100000
$QUERY cost claude-sonnet-4-6 --input 50000 --output 5000 --monthly 1000

# Top models by metric
$QUERY top --by cost          # cheapest accessible models
$QUERY top --by ctx --count 5 # largest context windows

# Provider overview
$QUERY providers

# Full record for a model
$QUERY get gemini-2.5-flash

---

Common scenarios

"What model should I use for [task]?"

sessions_spawn({
  agentId: "tool-models",
  runtime: "subagent",
  task: `recommend "${task}" — return top 3 with reasoning`
})

"How much will this cost?"

sessions_spawn({
  agentId: "tool-models",
  runtime: "subagent",
  task: `Estimate monthly cost for processing ${inputTokens} input + ${outputTokens} output tokens/day using ${model}. Also compare with the top 2 cheaper alternatives.`
})

"Is [model] available and does it support [capability]?"

sessions_spawn({
  agentId: "tool-models",
  runtime: "subagent",
  task: `get ${model} — check if it supports vision and PDF input, and what context/output limits apply on each provider`
})

"What's the best free model for [task]?"

sessions_spawn({
  agentId: "tool-models",
  runtime: "subagent",
  task: `recommend "${task}" --cost-max 0 — only free models`
})

---

Notes

• accessible = we have a working API key for that provider right now
• piaiAvailable = model is in OpenClaw's built-in catalog (subset of accessible)
• Provider-reported limits always override generic specs (SambaNova caps many models at 3k output; Groq caps at 32k — these differ from the model's stated max)
• Database refreshed nightly by the model-scanner cron job
• Data sources: OpenClaw pi-ai catalog → LiteLLM → OpenRouter → live provider APIs