Apify MCP Server

What it does

This server turns Actors from Apify Store into tools an assistant can call. Actors are ready-made scrapers and automations, say for Google Maps or Amazon. The server reads an Actor's input schema and builds a tool from it, so the assistant knows which arguments to send. Calling an Actor hands back the run status and storage IDs, not the scraped items. The assistant then pulls the items with get-dataset-items, paging with limit and offset.

Out of the box you get the actors and docs groups. The actors group covers Store search, Actor details and running an Actor, two tools for Actors that are MCP servers themselves, and send-server-actor-request, which sends one HTTP request to a Server Actor's API. On top of that come the Apify docs search and fetch, Web Search and Web Fetch as Actor tools, and four run and storage tools that join whenever call-actor or an Actor tool is loaded. report-problem sits in its own dev group in the code. The README says it is served by default only with telemetry on and a client whose name has none of claude, anthropic and local-agent-mode in it. More groups are opt-in through the tools URL parameter or the --tools flag: runs, storage, tasks, schedules, builds, versions and api. get-actor-list and delete-actor sit in no group and only load when named.

Counting src/tools/registry.ts at the published release gives 45 helper tools. With the two default Actor tools that makes 47, and there are four widget tools on top for MCP Apps clients. Only the search-actors and fetch-actor-details widgets get added on their own when UI mode is on.

Apify recommends the hosted endpoint at https://mcp.apify.com (Streamable HTTP, OAuth or an Authorization: Bearer header). Locally it runs over stdio with npx @apify/actors-mcp-server. Either way the Actors run on Apify's cloud and are billed to the account. The README also describes an experimental route without an account: AGI sells a prepaid, spend-capped Apify token paid for with x402, MPP or Skyfire.

What it needs

The local server takes an Apify API token from APIFY_TOKEN, or from ~/.apify/auth.json left by the Apify CLI login; it exits at start without one unless only public tools such as docs are selected. It also exits on Node.js older than 22 (src/stdio.ts). On the hosted endpoint, OAuth clients sign in through the browser; others send Authorization: Bearer <APIFY_TOKEN>.

Limits

• Cost Each Actor run bills to your Apify account at that Actor's own price. callOptions maxTotalChargeUsd or maxItems on call-actor caps it.
• Rate limit The hosted endpoint accepts up to 20 requests per second per account.
• Not covered search-actors only finds free and pay-per-event Actors that pass Apify's agent-safety filters, so rental and pay-per-result Actors are hidden.
• Limit Full-permission Actors need a one-time approval. Until then call-actor returns an error with an approval link.
• Limit The default tool set may change between versions; Apify advises naming tools explicitly for production use.
• Limit tools=source, add-actor, experimental, preview and get-actor-log are retired selectors and load no tools.
• Limit call-actor no longer takes the actorName:toolName form. A selection naming call-actor alone needs call-actor-mcp-tool added for MCP server Actors.
• Not covered Registers no prompts, so prompts/list comes back empty.
• Size cap resources/read inlines responses up to 256 KB.
• Early release Paying per run through AGI with x402, MPP or Skyfire is marked experimental.

Who it suits

Suits developers who want an assistant to pull data from websites through existing Apify scrapers, or to manage their own Apify Actors, builds, versions, tasks, schedules and storage. The hosted endpoint needs nothing installed and signs in with OAuth. The local stdio server needs Node.js 22 and an Apify token. It is official and can change and delete things on the account (Actor versions, schedules, Actors, raw API writes), so give it a narrow tools list where that matters. Since runs cost Apify credit, it suits less well someone who only needs to fetch an occasional page for free.