Proposal: Progressive tool discovery

[hemanth]

Problem

Tools are loaded into context at LanguageModel.create() time. On-device models have tight context windows (2K–8K tokens). Many tools with schemas eat into that fast.

Proposal

A toolSearch callback for on-demand tool discovery:

const session = await LanguageModel.create({
  tools: [/* always-on tools */],
  toolSearch: async (query) => {
    return registry.filter(t => matches(t, query));
  }
});

The UA exposes toolSearch as a built-in capability. Model calls it when it needs something not in its current tool set. Matched tools get added to the session dynamically.

Why

• 50 tools ≈ 5K+ tokens in schemas alone
• On-device models cannot afford that
• Only 2–3 tools are relevant per turn

Prior Art

• Cloudflare Code Mode — 2 meta-tools for 2,500+ endpoints
• Anthropic Tool Search
• WebMCP per-tool scopes