mcp.soubiran.dev is the MCP endpoint for my public content. It lets assistants search, query, and retrieve content from soubiran.dev, talks.soubiran.dev, and infra.soubiran.dev.

The service connects content that is already published by those websites. Each website remains the source of its own catalog and documents, while the MCP gives clients one interface across all three.

Why catalogs stay with each website

Each website generates a JSON catalog during its build. The catalog contains stable IDs, canonical URLs, and the metadata needed to discover its documents. The MCP fetches these catalogs and validates them before using their data.

This keeps publication and discovery tied together. A website deploys its content and catalog as one version, then the MCP reads that public version. Adding another content source requires a catalog adapter, a document retrieval rule, and a corresponding search index.

This approach fits public, read-only content spread across several static websites. It keeps the MCP focused on discovery and retrieval instead of turning it into another content store.

How the three tools work

The server exposes three tools:

  • search_content performs a semantic search with Cloudflare AI Search. It matches each result with the corresponding catalog entry, then returns an excerpt, category, score, and stable content ID.
  • get_content resolves one stable ID across the three catalogs and fetches the complete Markdown page or talk transcript from its source website.
  • list_content runs a read-only JavaScript query against the complete catalogs. It handles questions that are easier to express as code, such as listing recent posts or filtering talks by topic.

search_content and list_content return stable IDs that get_content can resolve. It rejects duplicate IDs across catalogs rather than choosing one result, so a new source cannot silently change which document an existing ID returns.

Code Mode without broad access

list_content follows the Code Mode pattern. The client sends one asynchronous JavaScript function, and the server runs it against frozen copies of the three catalogs in a temporary Worker.

The temporary Worker cannot access the network, secrets, storage, or bindings. It receives frozen catalogs and returns JSON to the MCP Worker. This boundary supports filtering and aggregation without giving generated code access to the MCP Worker or the source websites.

Catalog questions vary widely. One constrained query tool avoids a growing set of narrow tools and reduces the result before it enters the model context.

Deployment and observability

The MCP runs on Cloudflare Workers and is deployed with Wrangler. Cloudflare AI Search provides semantic results, while Worker Loader creates the isolated Workers used by list_content.

Cloudflare records logs and sampled traces. Sentry captures exceptions, and Evlog records tool outcomes, durations, selected upstream timings, and result sizes. Queries, content, excerpts, authorization headers, and cookies are redacted from those logs.

Ecosystem