Mohamed Eid, a developer behind the Arabic SEO resource 'SEO Bank,' has released a lightweight remote Model Context Protocol (MCP) server that allows AI agents to query his knowledge base directly. Built on Cloudflare Workers without any official SDKs, the server runs on a minimal codebase, demonstrating that robust AI integration doesn't require complex infrastructure. The project addresses a specific gap in the market: while English SEO advice is abundant, Arabic resources often suffer from thin, copied content. By exposing structured data via MCP, Eid enables tools like Claude and Cursor to provide accurate, cited answers to technical SEO questions.

Stateless Architecture and JSON-Only Responses

Eid made a deliberate architectural choice to ignore the Streamable HTTP transport's session and Server-Sent Events (SSE) capabilities, which are often overkill for read-only knowledge bases. The server operates as a stateless endpoint where every POST request is self-contained, returning plain application/json responses rather than streaming data. GET requests return a 405 Method Not Allowed, explicitly signaling the absence of server-to-client streams, while notifications receive a 202 Accepted status with no body. This design makes the endpoint trivially cacheable and allows it to coexist with the existing Astro-built website on a single Cloudflare Worker route, eliminating the need for separate maintenance or keep-alive processes.

Native Arabic Search Without Vector Databases

Handling Arabic language queries required specific normalization logic rather than relying on standard keyword matching, which fails due to the language's morphological complexity. The server strips diacritics and tatweel, unifies different forms of the letter alef, and removes common prefixes like 'the' or 'with the' before scoring relevance. Field weights prioritize titles over tags, summaries, and body text, ensuring high-precision results for just 86 well-structured documents. Eid notes that vector databases were unnecessary for this dataset size, proving that clever preprocessing can outperform heavier ML infrastructure for smaller, domain-specific corpora.

Optimizing Tool Definitions for AI Models

The developer refined his tool descriptions based on feedback from Glama, an MCP server registry that grades tool definitions. Initially scoring 16/20, Eid improved clarity by distinguishing between search_seo_questions (keyword-ranked) and list_seo_questions (chronological), adding pagination support with total and nextOffset fields. He also introduced get_seo_category to allow agents to browse specific sections. Crucially, all tools now include MCP annotations such as readOnlyHint: true and destructiveHint: false, giving client applications confidence that these operations are safe and idempotent. This attention to metadata ensures agents can trust the server's behavior without executing risky side effects.

Key Takeaways

  • The MCP server is implemented on Cloudflare Workers, bypassing official SDKs entirely.
  • Statelessness is achieved by rejecting SSE streams and using plain JSON responses, making the endpoint cacheable.
  • Arabic search accuracy is handled via manual normalization of diacritics and prefixes, avoiding vector DB overhead.
  • Tool definitions include strict MCP annotations to signal read-only safety to AI clients like Claude and Cursor.
  • The server is accessible via a single command: claude mcp add --transport http seo-bank https://seoquestionsbank.com/mcp.

The Bottom Line

Eid's project is a masterclass in restraint, proving that for static knowledge bases, a simple JSON-RPC switch is superior to bloated SDKs. It challenges the assumption that AI integration requires heavy infrastructure, showing that a well-structured, minimal codebase can effectively bridge the language gap in SEO tooling.