Model Context Protocol
Connect any AI agent to your personal knowledge base with a scoped, revocable token
A hosted MCP server gives Claude, Cursor, and any MCP-compatible agent access to your bookmarks — no binary to install.
What is MCP?
The Model Context Protocol is an open standard published by Anthropic that defines how AI agents request and receive structured context from external sources. It is the interface between a personal knowledge base and any AI assistant you choose to use.
MCP supports both local and hosted server modes. Blogmarks uses the hosted model: the MCP server runs on AWS Lambda and is reachable at api.blogmarks.dev. AI clients authenticate with a per-connection bearer token and receive only the bookmarks their scope allows.
The protocol is AI-provider-agnostic. Claude, Cursor, GPT, or whatever comes next — any agent that implements MCP can query your knowledge base without you changing anything on your end.
Why it matters
The default model for AI personalization is data upload: you paste content into a chat window, attach files to a session, or give a service permission to read your email. In every case, your personal data moves to the AI provider's servers to be processed.
MCP provides structured retrieval instead of bulk upload. The AI agent makes a specific query; the MCP server resolves it against your bookmarks; only the relevant results are returned as context. The AI sees what it asked for — not your entire reading history.
This matters because AI assistants are becoming infrastructure. Scoped, revocable tokens give you control over what each client can see, and activity logs let you audit every tool call.
How Blogmarks uses this
Blogmarks exposes your knowledge base through a hosted MCP server (Node.js Lambda) at mcp.blogmarks.dev/mcp. It is AI-provider-agnostic — Claude, Cursor, or any MCP-compatible agent can query what you have saved.
Each connection uses a unique scoped bearer token. Tokens are stored as one-way hashes server-side — the plaintext is shown once and never stored. Connections can be paused or revoked independently from the Settings → MCP Integrations page.
The MCP server is a read-only interface by default. Write access (the tag_bookmark tool) is opt-in per connection. The server enforces scope rules server-side regardless of client claims.
Further reading
- [1]
- [2]Abhinav Kimothi. A Simple Guide to Retrieval Augmented Generation. Manning Publications, 2025
- [3]Salvatore Raieli, Gabriele Iuculano. Building AI Agents with LLMs, RAG, and Knowledge Graphs. Packt Publishing, 2025