MCP integration

One config line.
Your bookmarks become context.

ContextBolt Pro gives your library its own MCP endpoint. Every page, post and thread you have saved becomes readable by Claude, Cursor, Windsurf, Cline and Zed mid-conversation, answered in plain English with the source links attached.

Get Pro See the six tools
Pro $6/mo · unlimited bookmarks · the endpoint is Pro only
Claude Desktop your library

Click to try a real question

Your own endpoint, answering from your own saves. Every result links back to the post it came from.

On this page
What is MCP Tools exposed Client setup Example prompts Privacy and security
001 / The protocol

What MCP is,
plainly.

Model Context Protocol is an open standard, originally from Anthropic, that lets AI agents talk to external tools and data in a uniform way. Think of it as USB for AI: one connector, many devices.

Before MCP, each tool wrote a bespoke integration for each agent. With MCP, you write one server and every client speaks to it. ContextBolt is one such server, for your social bookmarks.

Per query

What happens when you ask.

  1. 01 You ask“What did I save about rate limiting?” Claude does not need to know we exist.
  2. 02 The agent calls a toolIt sees search_bookmarks registered over MCP and calls your own endpoint, over JSON-RPC.
  3. 03 A cited answerWe search your index and return the top matches. Claude composes the reply, and you see the source links.
002 / Tools exposed

Six functions.
That's it.

A small surface on purpose. Less to learn, less to break, and each one maps to something you would actually ask for. You never call these yourself: you ask a question and the agent picks.

Every endpoint publishes its own schema at tools/list, so your agent reads the parameters on connect and you never type one. The tool reference shows what to ask for each.

search_bookmarks
most used

Semantic search across the entire library. Returns the top matches ranked by cosine similarity over 1536-dimension embeddings.

list_clusters

List every topic cluster discovered in your library. Useful when an agent wants to take stock before drilling in.

get_cluster_bookmarks

Fetch every bookmark inside one cluster. Pairs naturally with list_clusters for "show me everything about X".

get_recent_bookmarks

Recent saves, in reverse chronological order. Used in morning briefings and "catch me up since last week" prompts.

save_bookmark
write

Save a new bookmark to your library from inside an agent. ContextBolt tags and embeds it server-side, so it is searchable straight away.

export_bookmarks

Hand back a copy of the collection so it can be backed up, migrated or used somewhere else. Your library is never locked in.

003 / Client setup

One line.
Any MCP client.

Drop your endpoint URL into your client's MCP config and restart. Your token is generated in the extension, under Settings, MCP.

Three shown here, and the setup guide covers Claude Code and Cline. Anything else that speaks MCP takes the same URL; the key is what differs, and Zed's is one of the three below.

Claude Desktop
macOS / Win / Linux 
// claude_desktop_config.json
{
  "mcpServers": {
    "bookmarks": {
      "url": ".../mcp/<token>"
    }
  }
}
after editquit and relaunch
Zed and any host
JSON-RPC 
// settings.json
// context_servers, not mcpServers
{
  "bookmarks": {
    "url": ".../mcp/<token>"
  }
}
specmodelcontextprotocol.io
Cursor
IDE 
// .cursor/mcp.json
{
  "mcpServers": {
    "bookmarks": {
      "url": ".../mcp/<token>"
    }
  }
}
verifySettings, MCP
004 / Prompts that work

What to ask
your agent.

You don't have to mention ContextBolt. Just ask the kind of question that only your library could answer. The agent figures out which tool to call.

What did I save about rate limiting last month?
Calls search_bookmarks with the question and a time bias. Returns ranked posts with similarity scores.
Which topics am I obsessed with right now?
Calls list_clusters. Returns top topics by recent activity. Surprisingly revealing.
Give me the weekly catch-up: what should I read from last week?
Calls get_recent_bookmarks with a 7-day window, filtered by cluster relevance.
Draft a thread based on what I saved about MCP.
Calls search_bookmarks, then generates. Your saved points of view become your published ones.
005 / Privacy and security

Your library
stays yours.

The library itself lives in IndexedDB on your device. Pro syncs a copy to our servers so the endpoint has something to read, and that copy is readable by us. It is not end-to-end encrypted. Basic never syncs, so your library is not stored with us. Two things still do reach us. The text of a save is sent up to be tagged and embedded. A bulk import bigger than your visible 150 is the exception, because the overflow is parked unprocessed, but anything you save one at a time goes through. And installing creates an anonymous account, which the extension checks in against every few minutes while your browser is open, logging your IP each time. The privacy policy has the full list.

Local first

Your library and its search index live on your device, and the extension searches them there. The endpoint is the other half: it reads the synced copy on our servers, so anything your AI tool asks for is a round-trip.

Rotatable tokens

Your MCP token can be rotated from the extension at any time. Old configs stop working straight away, no support ticket needed.

No query retention

We log query counts for billing, never query content. Usage is readable in the extension dashboard.

Personal MCP endpoint, $6 a month

Plug in your bookmarks.
See your agents change.

Pro takes about ninety seconds to set up. Most of that is restarting your client.

Get Pro Read the setup guide