DataForSEO Brain

MCP Server Integration

4 min read

MCP Server Integration

The official DataForSEO Model Context Protocol server that lets AI assistants call SEO APIs through natural language. Sits under DataForSEO Brain → Concepts.

Overview

The official DataForSEO MCP server is a TypeScript/Node.js implementation of Anthropic’s Model Context Protocol, published under Apache-2.0 (around 221 stars / 120 forks on GitHub in mid-2026). It exposes selected DataForSEO APIs as MCP tools so agents like Claude Code, Cursor, and ChatGPT can request SEO data via prompts instead of hand-coded REST calls. It uses a JSON-RPC client-server architecture and removes per-source integration work. The package is published on npm as dataforseo-mcp-server (latest 2.9.9 at retrieval, Jun 2026) and requires Node.js >=20.0.0 (per package.json engines; the README still says v14+ and the marketing page 18+, both stale). The v2.9.9 line added Historical SERPs (date params), an exclude_targets field on the backlinks domain-intersection tool, tool annotations/titles, OAuth/MCP auth improvements, and Cloudflare Worker deploy. The vendor and repo are profiled in ent-dataforseo-mcp-server.

What it covers

The server is organized into modules, each wrapping an API family and exposing multiple tools. The README (the authoritative superset) lists ten modules:

  • SERP (Google, Bing, Yahoo results), KEYWORDS_DATA (volume, CPC, clickstream), ONPAGE (crawling), DATAFORSEO_LABS (proprietary databases), BACKLINKS (links and referring domains), BUSINESS_DATA (reviews, listings), DOMAIN_ANALYTICS (technologies, Whois), CONTENT_ANALYSIS (citation tracking, sentiment), AI_OPTIMIZATION (keyword discovery, conversational optimization, LLM benchmarking), and MERCHANT (product, pricing, seller data).
  • Marketing pages sometimes list only the “seven major” modules and omit AI_OPTIMIZATION and MERCHANT; this is a documentation lag, not a contradiction.
  • By default the server registers hundreds of tools, which consumes significant LLM context, so operators are expected to narrow the surface.
  • Officially documented clients include Claude Desktop, Claude Code, Cursor, ChatGPT, Gemini CLI, and Docker; reported agent workflows include keyword clustering, backlink comparison, and competitor analysis from conversational prompts.

Key parameters / inputs

Environment variables and config:

varrole
DATAFORSEO_USERNAMErequired; API login (account email)
DATAFORSEO_PASSWORDrequired; auto-generated API password (not the dashboard password)
ENABLED_MODULEScomma-separated allow-list (for example “SERP,KEYWORDS_DATA,ONPAGE”); primary permission gate; all modules load if unset
ENABLED_PROMPTSallow-list of prompt templates (for example top_3_google_result_domains)
DATAFORSEO_FULL_RESPONSEtrue returns unfiltered raw API responses (default false)
DATAFORSEO_SIMPLE_FILTERtrue flattens the filter schema for LLMs that struggle with nested structures

A field-config.json (supported_fields keyed by tool name) further trims returned fields via --configuration field-config.json.

Response / what you get back

Tools return the underlying DataForSEO API result, optionally filtered. Modules are built on a BaseModule class (a getTools() method) registered in modules.config.ts. Transports available:

  • stdio (default): npx dataforseo-mcp-server, the local mode Claude Desktop uses.
  • Streamable HTTP: npx dataforseo-mcp-server http (port 3000); works with most modern MCP clients but not Claude Desktop.
  • SSE (legacy/deprecated): npx dataforseo-mcp-server sse.
  • Hosted remote: DataForSEO runs managed endpoints at https://mcp.dataforseo.com/mcp (streamable HTTP) and https://mcp.dataforseo.com/http (HTTP/SSE), so no self-hosting is required.
  • Cloudflare Worker: an index-worker.js build with POST /mcp, GET /sse, GET /health, deployable via wrangler.

Cost & method notes

  • The MCP server itself is free and open source; usage bills against your DataForSEO account at the same per-endpoint cost as raw REST. See cap-queue-priority-cost-model.
  • Curated MCP can cut LLM tokens roughly 50-80% with no consumer-side integration code, at the cost of roughly 15-25% added latency from JSON-RPC overhead; naive MCP with too many verbose tools can increase tokens, which is exactly why ENABLED_MODULES and field-config exist.

When to use / how it fits

Choose MCP for agent-native prototyping and interactive runtimes (Claude Code, Cursor, ChatGPT), multi-tool reasoning, and going from idea to running prototype in minutes. Choose raw REST or the official client libraries for production batch jobs, scheduled enrichment, and high-throughput statically-authenticated server-to-server calls. The full tradeoff is the decision dec-mcp-vs-raw-rest. Auth is the same HTTP Basic scheme as REST, covered in cap-authentication-security. The AI_OPTIMIZATION module surfaces the capabilities in cap-ai-optimization-api.

Gotchas / limits

  • Default loads all modules and hundreds of tools, drowning the agent’s context; always set ENABLED_MODULES.
  • MCP suits interactive sessions; many MCP paths assume a per-user session and lack a clean static-credential server-to-server option, where REST is stronger.
  • Streamable HTTP transport does not work with Claude Desktop; use stdio there.
  • The same 2000 calls/min and per-endpoint limits in cap-rate-limits-throughput apply, since MCP is a thin wrapper over the same APIs.

Sources

Graph View

Backlinks