What it takes off you is the build. The store search looks for an Actor that already handles the site, the details call returns the input it expects, and the dataset call reads the rows out — so scraping a site becomes a lookup and a run rather than a project.
The Apify platform as MCP tools. Actors are searched, inspected for their input schema, called, and their runs and outputs read — with the Apify and Crawlee documentation searchable from the same server. Around ten tools load by default; twenty-five exist, selectable by category or by name.
- The Apify Store searched by keyword, and an Actor's details fetched — input schema, README, pricing and output schema — before it is run (`search-actors`, `fetch-actor-details`)
- Any Actor called with input built from that schema, its run inspected, and a running one aborted (`call-actor`, `get-actor-run`, `abort-actor-run`)
- Dataset items read with filtering and pagination, and a single key-value store record fetched by key — auto-injected alongside the calling tools, because a call returns run metadata and storage IDs rather than the rows themselves (`get-dataset-items`, `get-key-value-store-record`)
- Two web Actors in the default set: one that browses the web for retrieval, one that fetches a URL and returns its content (`apify--rag-web-browser`, `apify--web-fetch`)
- Apify and Crawlee documentation searched by full text and fetched by URL (`search-apify-docs`, `fetch-apify-docs`)
- Off by default and available by category: run lists and logs, dataset and key-value store metadata, schema generation and key listing (`get-actor-run-list`, `get-actor-log`, `get-dataset-schema`, `get-key-value-store-keys`)
- Saved Actor tasks — named, reusable configurations — created, read, updated, and published to or withdrawn from a public landing page (`create-actor-task`, `update-actor-task`, `publish-actor-task`)
- Any Apify API GET URL read on demand through the resources interface, with the session's token injected; responses inline up to 256 KB and return a download URL above that
- A problem reported back to the Apify team from inside the session (`report-problem`)
The hosted server at `https://mcp.apify.com/` with an Apify API token as `Authorization: Bearer <token>`, or `npx @apify/actors-mcp-server` locally. Which tools load is set by `tools=` — as a query parameter on the hosted URL or a `--tools` flag on the CLI — and specifying it replaces the defaults rather than extending them; `tools=apify/my-actor` exposes exactly one Actor and nothing else. An autonomous agent with no Apify account can pay per run instead: AGI issues a prepaid, spend-capped token, direct x402 signs USDC on Base and refunds the unused balance after 60 minutes idle, and Skyfire uses a PAY token of at least $5.00. Discovery calls are free under all of them. A payment-only session cannot use the API-read resources, which still need a token.
One command plus a key — npx @apify/actors-mcp-server, then supply credentials