MCP, both directions.

As a client

Plug your CRM, KB, or internal tools into the agent in one line.

import { mcpClientMount } from "@linda/core";

linda.mount("/host/crm",
  mcpClientMount({
    transport: "sse",
    url: "wss://crm.example.com/mcp",
    auth: { type: "bearer", token: process.env.CRM_TOKEN },
  })
);

// The LLM can now read /host/crm/accounts/123.json, call /host/crm/tools/search, etc.

As a server

Expose your page's VFS + tool primitives over MCP so Claude Desktop, Cursor, or Continue can read it directly.

import { LindaMcpServer } from "@linda/core";

const server = new LindaMcpServer({
  linda,
  expose: {
    resources: ["/page", "/conversation", "/user"],
    tools: ["read", "write", "act", "extract", "observe"],
  },
});

// Connect from Claude Desktop:
// { "mcpServers": { "linda": { "transport": "sse", "url": "wss://..." } } }

Plug the URL into Claude Desktop's mcpServers config and the agent in Claude Desktop can read /page, /conversation, and call act / extract / observe on your page from outside the browser.

Why MCP matters here

Patterns

• Agent + your KB. Mount your Notion / Confluence MCP under /host/kb. The agent grounds answers without retrieval boilerplate.
• Agent + your CRM. Mount Salesforce / HubSpot MCP. The agent enriches form fills with account context.
• Claude Desktop + your page. Expose Linda as MCP. Your PM/sales team can talk to Claude Desktop and have it read the same page the user is on, in real time.

Read the full MCP docs →