Apify published a walkthrough of using MCP with its Actors, and the pitch is that one connector replaces a pile of custom integrations. Add https://mcp.apify.com to Claude as a custom connector, sign in through the browser, and ask for something like “a list of 10 cafes in Boston with at least 4 stars on Google Maps.” Apify says Claude picked Google Maps Scraper from Apify Store on its own and returned results within 15 seconds. No code, no install, no pasted token.
“Thousands of public Actors” and “the largest marketplace of tools for AI” are Apify’s own framing, and the 15-second run is one example from the vendor. What survives is the mapping, which is the most useful part of the post. Actors are the tools. Datasets and Key-Value Stores are the resources, reached through tools like get-dataset-items and get-key-value-store-record. The other line worth keeping is a quiet one: you can set the tools to Needs approval, because “runs use your Apify credits, you stay in control of spending.” An agent that chooses its own Actor is also choosing what you pay, and here the brake lives in the client’s connector settings rather than in the server.
The catalog lines up with that architecture. Apify has six API pages, and the story touches four: the Apify Actors API behind discovery, the Apify Actor Runs API behind execution, and the Apify Datasets API and Apify Key-Value Stores API behind the results. The OAuth sign-in checks out too, with delegated identity, protected resource metadata and dynamic client registration all lit, which is the machinery that lets a browser sign-in replace a pasted token. The agentic access profile records 13 operations, three of them acting.
The measurement splits in an unusual direction. Agent Readiness is 52.9, agent-ready, ahead of a Kin Score of 35.5, thin. Discoverability is 80.0 and developer ergonomics 62.4, while operational transparency is 10.5, contract governance 13.6 and contract quality 26.6. Apify is easier for an agent to connect to than its contracts are to rely on. The unlit dimensions are the ones this story raises: idempotency, error semantics, reversibility and dry run. Those are what would let an agent retry a paid run safely, or preview one before spending anything. The post also calls MCP a foundational layer of the agentic commerce stack, and that dimension is unlit as well. For now, the approval button is doing the work those dimensions would do.