Best MCP Server Actors 2026

Contains affiliate links. Disclosure.

The Model Context Protocol (MCP) is an open standard (originated by Anthropic, now community-governed) for handing an LLM a toolbox. Think of it as a USB-C port for AI: one connector, many tools. Claude (Desktop and Code), Cursor, Windsurf, ChatGPT, Codex and VS Code all speak it, and Apify publishes a hosted MCP server at mcp.apify.com that fronts the entire Actor catalog. If you want to test these tools against the flagship Claude models, you can try Claude free for a week.

Most Actors can technically be called through MCP. Few should be. Good MCP tools return small, structured JSON with a tight schema, not 40-page markdown dumps that eat your context window. This list is the short set that actually earns its place in an agent config.

The Apify Store lists 30,000+ Actors and a growing MCP category, but for day-to-day agent work these three cover most of the job.

• Step-by-step setup: wire MCP into Claude Desktop, Cursor, and Windsurf.

Top MCP Actors comparison

Tool	Best used as	Payload shape
Apify MCP Server	Universal hosted MCP endpoint	Dispatches to any Actor
RAG Web Browser	Dedicated MCP tool for live search	Small markdown chunks
Website Content Crawler	Batch RAG ingestion, not live agent use	Large markdown corpus

Check each Actor's live user count and rating on its Apify Store page before committing to it.

---

1. Apify MCP Server

Apify runs the MCP server as a hosted Streamable HTTP endpoint at https://mcp.apify.com/. You don't install anything: you paste the URL into your client and authenticate. By default it enables tools for Actor discovery, Apify docs search, and the RAG Web Browser Actor, plus any specific Actors you pin.

• Transport: Streamable HTTP is the recommended path (Apify retired the older HTTP+SSE transport in 2026). The @apify/actors-mcp-server npm package is still available as a local stdio server for clients that don't speak remote MCP, but the hosted endpoint is what you want for Claude Desktop, Cursor, ChatGPT, and Codex.
• Auth: OAuth (Claude Desktop, Claude.ai, ChatGPT) or a Bearer token (APIFY_TOKEN from console.apify.com/settings/integrations) for manual configs.
• Tool selection matters: the default config loads the generic Actor-discovery tools. For a focused agent, pass an explicit tool list via the tools query parameter (for example tools=apify/rag-web-browser,apify/google-search-scraper) so the LLM sees a few clear schemas instead of a search tool over 30,000+ Actors.
• Pricing: the server itself is free; you pay per Actor run (same compute units as running the Actor directly).

When to reach for it: you want your agent to have open-ended web research ability, and you're comfortable with it browsing Store and calling whatever fits. Pair with pinned Actors below for reliability.

2. RAG Web Browser

RAG Web Browser is built for this protocol. You give it a query or URL, it runs a Google search and/or fetches the page, strips chrome, and returns a small markdown payload plus metadata. Average response is a few KB per result, so it fits inside a tool-call turn without torching your context.

• Pin this one: it's enabled by default in Apify's own hosted MCP setup (apify/rag-web-browser).
• Why it works as an MCP tool: bounded output, one clear input (query), predictable runs, and it's a close replacement for Firecrawl's search and scrape tools.
• vs Firecrawl MCP: comparable quality on markdown extraction; you pay per result on Apify's compute model instead of Firecrawl's credit system, and you stay inside one vendor if you're already using Apify for heavier scraping.

3. Website Content Crawler

A strong choice for full-site ingestion: sitemap-aware, handles JS rendering, deduplicates, outputs clean markdown. It is one of the most-used and highest-rated crawlers in the Store, which is why it ships in most Apify RAG tutorials.

It's also the Actor most likely to wreck an agent conversation if you expose it raw. A crawl of a mid-size docs site returns tens of thousands of tokens. Guidance:

• Prefer batch mode: run it via the API or schedule, push output to a vector store, and let the agent query the store. That's the standard RAG pattern and MCP isn't the right layer here.
• If you must expose it as a tool: hard-cap maxCrawlPages (5–20) and maxResults, and consider wrapping it in a custom MCP server that paginates or summarizes before returning.

---

How the pieces fit

Claude / Cursor / Windsurf
        │  (MCP over Streamable HTTP)
        ▼
https://mcp.apify.com/   ──►  actors tool  ──►  any Actor in Store
                         ──►  docs tool    ──►  Apify docs search
                         ──►  pinned Actors (e.g. rag-web-browser) ──► direct call

One connection, many tools. The pinned Actors get their real input schemas surfaced to the LLM, which is what makes them reliable to call. The generic actors tool is for the long tail.

Picking Actors that work well over MCP

Not every Actor is a good MCP citizen. Checklist:

1. Small, structured output. JSON with a handful of fields beats a 50KB HTML-to-markdown dump every time.
2. Fast runs. Under ~30s so the model doesn't stall. Cheerio-based scrapers over Playwright when both work.
3. Narrow inputs. A query string or a single URL is ideal. Actors with 20 config fields confuse the tool-call path.
4. Idempotent-ish. Agents retry. Actors that bill heavily per call should be gated or replaced with cached wrappers.

Good candidates beyond this list: Google Search Scraper for SERPs, and any of the social/profile scrapers when you want a specific entity rather than a crawl. Full-site crawlers and anything that paginates into thousands of items belong out of the agent loop.

Use cases

• Research agents: pin RAG Web Browser; the agent searches and reads pages inside the conversation.
• Dev-tool agents in Cursor / Windsurf: pin docs + a handful of scrapers so the agent can answer "how do I call X Actor" and run it in the same turn.
• Monitoring agents: pin a specific scraper (prices, mentions, jobs) and schedule the agent prompt.

Connect your agent

Paste https://mcp.apify.com/ into Claude, Cursor, or ChatGPT and authenticate.

Open the Apify MCP Server →

Frequently Asked Questions

Use the hosted Streamable HTTP endpoint at mcp.apify.com for anything modern: Claude Desktop, Claude.ai, Cursor, ChatGPT, Codex, VS Code. Use the @apify/actors-mcp-server npm package as a local stdio server only if your client doesn't support remote MCP, or if you need to run the server behind a firewall.

Two options. OAuth 2.0 for Claude Desktop, Claude.ai, and ChatGPT: you click through once and the client stores the token. Or Bearer auth with your Apify API token from console.apify.com/settings/integrations for manual JSON configs and custom clients.

The MCP server is free. You pay only for the Actor runs it triggers, billed the same as running those Actors directly. The $5 monthly free credit on the Apify free plan covers light agent use.

Firecrawl MCP exposes a focused set of tools (search, scrape, crawl, extract) over their own infrastructure. Apify MCP exposes the entire Actor ecosystem (30,000+ Actors as tools) with RAG Web Browser as the direct Firecrawl-equivalent. If you only need search and scrape, both work; if you also need Instagram, Google Maps, LinkedIn, Amazon, or custom scrapers, Apify is the one vendor that covers all of it.

Pin a small explicit tool list instead of using defaults. On mcp.apify.com use the Select Actors panel, or pass the `tools` query parameter in your client config (e.g. `tools=apify/rag-web-browser,apify/google-search-scraper`). Three to five tools with clear schemas beats dynamic discovery over 30,000+ Actors.