Model Context Protocol is the open standard for giving agents access to live data. ContextBolt Pro ships a personal MCP endpoint that exposes your library to any compatible client: Claude, Cursor, Windsurf, Cline, Zed, anything that speaks MCP.
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.
You ask Claude "what did I save about rate limiting?". Claude does not need to know we exist. It just sees a search_bookmarks tool registered via MCP and calls it.
The query goes to your ContextBolt endpoint. We search the index, run the embeddings, and return the top results. Claude composes a reply. You see the citations.
"What did I save about rate limiting?"
Claude sees the search_bookmarks tool and decides it is relevant.
JSON-RPC over HTTP to your /mcp/<token> endpoint.
Top matches returned with source links. Claude composes with citations.
We deliberately ship a small surface area. Less to learn, less to break. Each tool maps to one obvious thing an agent needs from a personal library.
The full schema is published in tools/list at every endpoint, so Claude introspects it on connect.
Semantic search across the entire library. Returns the top matches ranked by cosine similarity over 1536-dimension embeddings.
List every topic cluster discovered in your library. Useful when an agent wants to take stock before drilling in.
Fetch every bookmark inside one cluster. Pairs naturally with list_clusters for "show me everything about X".
Recent saves, in reverse chronological order. Used in morning briefings and "catch me up since last week" prompts.
Save a new bookmark to your library from inside an agent. ContextBolt tags and embeds it server-side, so it is searchable straight away.
Drop your endpoint URL into your client's MCP config and restart. Your token is generated in the extension, under Settings, MCP.
If your client speaks MCP but isn't listed below, the same URL works. It's just JSON.
// claude_desktop_config.json { "mcpServers": { "bookmarks": { "url": ".../mcp/<token>" } } }
// .mcp.json at project root // or ~/.claude/mcp.json globally { "mcpServers": { "bookmarks": { "url": ".../mcp/<token>" } } }
// .cursor/mcp.json { "mcpServers": { "bookmarks": { "url": ".../mcp/<token>" } } }
// ~/.codeium/windsurf/ // mcp_config.json { "mcpServers": { "bookmarks": { "url": ".../mcp/<token>" } } }
// Settings, MCP Servers // the Cline UI or json directly { "bookmarks": { "url": ".../mcp/<token>" } }
// settings.json // context_servers { "bookmarks": { "url": ".../mcp/<token>" } }
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.
search_bookmarks with the question and a time bias. Returns ranked posts with similarity scores.search_bookmarks, then composes a synthesis. Source links are preserved as citations.list_clusters. Returns top topics by recent activity. Surprisingly revealing.get_recent_bookmarks with a 7-day window, filtered by cluster relevance.search_bookmarks, then generates. Your saved points of view become your published ones.get_cluster_bookmarks, then reasons across them. Multi-source pattern matching.The MCP endpoint is the only thing on our servers. The library itself lives in IndexedDB on your device. Pro users opt into encrypted sync, and the keys never leave the browser.
Bookmark storage, embeddings, and the search index all live on your device. No server round-trips for reads.
Your MCP token can be rotated from the extension at any time. Old configs stop working straight away, no support ticket needed.
We log query counts for billing, never query content. Usage is readable in the extension dashboard.
Pro takes about ninety seconds to set up. Most of that is restarting your client.