/ writing / automation

AI Tooling Capability Union

2026-08-23 · cynix · 3 min read

The "what can my agent call right now" feed. This actor merges the canonical MCP server registry (Tier 1) and the OpenRouter model catalog (Tier 2) into a single normalized dataset under one shared capability taxonomy — so a single query returns the servers and the models that satisfy it, cross-layer. Nobody on Apify does this.

It's the Tier-3 play in the first-mover suite. Tier 1 indexed the MCP registry with an agentic matcher. Tier 2 indexed OpenRouter models with change detection. The Union is the combination: one capability-queryable surface over both tool layers, plus a per-capability coverage index that tells an agent exactly what's available.

What it does

Fetches both layers — MCP servers from registry.modelcontextprotocol.io (cursor-paginated, deduped by name) and models from openrouter.ai/api/v1/models (keyless JSON, 422+ models).
Normalizes each into one record shape with a shared capabilities tag set, so an agent reasons over servers and models identically.
Matches an optional plain-English taskQuery across both layers — "a vision model that can call a web search tool" surfaces vision models and vision-capable MCP servers together.
Emits a capability index — per capability tag, how many MCP servers vs models support it.

Input

{
  "taskQuery": "a vision model that can call a web search tool",
  "capabilities": ["vision"],
  "includeMcp": true,
  "includeModels": true,
  "maxMcpServers": 300,
  "capabilityIndex": true
}

Output (verified run)

A real run returned 302 MCP servers + 422 models merged, with 536 task matches and a capability index showing vision: mcp=5, model=59. Top matches for the query above included Exa web search (MCP, 100), Qwen reasoning models (100), and Nex-N2-Mini (multimodal, 100) — all in one feed.

Source quirks worth knowing

OpenRouter's public API does not expose modality flags for most models (0/422 carry an image modality), so vision/audio capability for models is derived from description text — the same way MCP server tags are. The shared taxonomy normalizes image→vision so a vision filter hits both layers. MCP pagination is cursor-based (?cursor=); ?offset= is ignored.

Why publishable, why first-mover

Reads only publicly published registry/model metadata — no personal data, no login, no ToS conflict. The Store has MCP-index actors and OpenRouter-index actors, but none union them under one queryable capability model. This is the novel layer.

AI Tooling Capability Union on Apify. Pay-per-event: a small per-run start fee plus a per-record item fee. Both sources keyless, no proxy cost.

← back to writing