Apify published a build for an AI research analyst in Slack, wiring n8n to their MCP server at mcp.apify.com so a question asked in a channel becomes a scrape run and a synthesised answer in the thread. The interesting mechanic is tool discovery rather than tool use: “an agent connected to the Apify MCP server can go further. It can identify the target platform, select a specialized Actor, and extract reviews or conversations as structured data.” The agent is not handed the RAG Web Browser or the Google Search Results Scraper up front — it finds the Actor that matches the question. That is the argument for an MCP server over a fixed toolset, and Apify is unusually well placed to make it, because a catalogue of thousands of Actors is exactly the case where enumerating tools in a prompt stops working.
One line in the setup is worth pulling out, because it is presented as a convenience and is really an architecture: “teammates don’t need Apify or OpenAI accounts—the workflow uses your API credentials, so they only need Slack access.” That is true, and it means every question anyone in the channel asks spends one person’s credits under one person’s identity. There is no per-user attribution, no per-user budget, and no way to revoke one colleague without revoking the workflow. For a research bot on a small team that is a reasonable trade made knowingly. It is also the exact shape that becomes a problem at the point it stops being a small team, and the tutorial would be stronger for saying so.
The catalog reads Apify as six API pages, and the tutorial exercises most of the chain. The Apify Actors API is what the MCP server is discovering over, the Actor Runs API is each invocation the agent triggers, the Datasets API is where the structured output lands before it is summarised back into Slack, and the Users API is where the credit balance being spent actually lives. Its agentic access profile counts 13 operations with 3 acting.
Apify scores 36.5, thin on the Kin Score — discoverability at 81.5 and developer ergonomics at 63.1 well ahead of contract quality at 27.3 and operational transparency at 10.5. Agent Readiness is 31.9, agent-ready, and mcp_server is lit, which is the catalog confirming the thing the post is built on rather than arguing with it. The unlit column is where the shared-key problem shows up as measurement: delegated_identity and consent_identity are both dark, which is the machine-readable version of “the workflow uses your API credentials.” So is idempotency — and on a platform where every Actor run costs credits, a retried question is a charged question. The MCP server is the finished half of this story. Telling a provider which human an agent is asking for is the half nobody has shipped.