How to Use Apify Actors with MCP Servers
Quick Answer
Use the Apify MCP server to connect Apify Actors to Claude Desktop, Cursor, and other AI clients. The MCP server exposes Actors as tools that the model can invoke during a conversation, so the assistant can start runs, pass inputs, and read results without you manually clicking through the Console.
Under the hood you still pay for Apify compute and Actor usage; MCP is a protocol layer on top of the same platform you use from the API or SDK.
What is MCP?
Model Context Protocol (MCP) is an open standard for connecting assistants to tools and data sources. Instead of pasting API docs into a chat, the client discovers tool schemas (name, description, parameters) and the model decides when to call them.
For Apify, each exposed Actor becomes a tool: the assistant can propose a run with structured input, wait for completion, and summarize or transform dataset output. With 30,000+ Actors in the Apify Store, this turns a single MCP connection into a broad catalog of web-data tools your assistant can reach for on demand.
MCP is an open standard (originated by Anthropic and now community-governed) supported across many clients, including Claude Desktop and Claude Code, ChatGPT, VS Code, and Cursor. If you do not have a client yet, you can try Claude free for a week and connect it to Apify in a few minutes.
Hosted Apify MCP (recommended)
The hosted endpoint is https://mcp.apify.com. It is the simplest path for Claude Desktop, Cursor, and other MCP-aware clients.
Prerequisites
- An Apify account
- An API token from Console → Integrations
- A client that supports MCP (and, for the hosted server, Streamable HTTP as documented by Apify)
Setup (high level)
- Add the MCP server as a remote connector in your client, pointing at
https://mcp.apify.com. In Claude Desktop you can search for Apify in the connector directory for a one-click install; Cursor and VS Code offer a one-click install button from the Apify MCP configurator. - Authorize via OAuth when prompted (you are redirected to your browser to sign in to Apify on first connect), or paste an API token for manual or local setups.
- Pick Actors to expose so the tool list stays small and relevant.
- Prompt in natural language; the assistant selects tools and parameters.
Transport note: The hosted server uses Streamable HTTP, in line with the official MCP specification. The legacy SSE transport was retired on April 1, 2026, so update any older client that still expects SSE; stale transports fail with unclear errors.
Authoritative steps and JSON snippets: Apify MCP documentation and the actors-mcp-server Store page.
Local MCP server (npx)
For development, air-gapped workflows, or advanced customization, run the open-source package. Pass your token via the APIFY_TOKEN environment variable and name the specific Actors you want to expose with --actors:
APIFY_TOKEN=your_token npx -y @apify/actors-mcp-server --actors apify/rag-web-browser
Keep the exposed set to one or two Actors. Loading the full catalog as tools bloats the model's context window and makes tool selection less reliable. You can list several Actors comma-separated if you need more than one.
The server process runs on your machine (or your VPC) instead of hitting only the hosted gateway, which is useful when you need tighter control over versions, logging, or corporate network policy. Details and the full flag list are in the Apify MCP docs linked above.
Example use cases
| Use case | What you might ask the assistant | Why MCP helps |
|---|---|---|
| Lead research | “Given this list of domains, find homepage titles and contact pages.” | Chains crawl/scrape tools without writing glue scripts in the chat |
| Maps / local data | “Pull top-rated cafes in Austin from a Maps Actor and tabulate ratings.” | Structured tool args instead of manual input JSON |
| Monitoring | “Run my saved Actor with these inputs and tell me if any field crossed a threshold.” | Combines execution + natural-language interpretation |
| RAG prep | “Crawl these URLs and summarize what changed vs. last week.” | Connects crawling Actors to your analysis step in one thread |
MCP vs. REST API
| MCP (with Apify) | Apify REST API / SDK | |
|---|---|---|
| Who orchestrates | The AI client proposes tool calls; the model picks tools and arguments | Your code or workflow builder orchestrates runs explicitly |
| Discovery | Tools advertised with schemas to the model | You read OpenAPI / SDK docs and write requests |
| Best for | Interactive research, ad-hoc extraction, copilots | Production pipelines, CI, deterministic ETL, strict SLAs |
| Control | Higher variance (model may mis-specify inputs) | Full control over inputs, retries, and idempotency |
| Security | Treat tokens like API keys; limit exposed Actors | Same token hygiene; easier to lock to one Actor ID in code |
Rule of thumb: Use MCP when a human is in the loop and speed of iteration matters. Use the API or SDK when you need repeatable jobs, tests, and deployment pipelines.
Cost and limits
- There is no separate MCP subscription; you pay for Actor runs, compute, and proxies like any other Apify usage.
- The free plan includes $5/month in platform credit (Apify free plan).
- Hosted MCP may enforce per-user request limits; heavy automation should move to scheduled Actors or the API.
Open Apify, grab your token, then follow the official MCP setup guide for your client. Start with one or two Actors exposed as tools to keep prompts reliable.
MCP is an open standard that lets AI assistants use external tools with structured parameters. Apify’s MCP integration exposes Actors as those tools so assistants can run scrapers and return dataset results during a conversation.
Use the hosted server at https://mcp.apify.com for most clients. For local development, run the @apify/actors-mcp-server package via npx as documented on Apify.
OAuth is recommended when your client supports the hosted flow. API tokens work for manual configuration or local servers; treat them as secrets and rotate them if leaked.
No additional MCP fee. You pay normal Apify usage for Actor runs and platform resources. Monitor spend in the Apify Console.
Any MCP-compatible assistant or IDE integration can work. Common setups include Claude Desktop, Cursor, and VS Code with MCP extensions; verify Streamable HTTP support against Apify’s current documentation.
Yes. Private or public Actors you own can be exposed as tools alongside Store Actors, subject to your MCP configuration and token permissions.
Security depends on token handling and which Actors you enable. Use least privilege, avoid sharing configs in public repos, and disable tools you do not need to reduce accidental high-volume runs.
Use the API or SDK for production automation, strict input validation, scheduled jobs, and testing. Use MCP for interactive assistant-driven workflows where the model helps choose and parameterize tools.
