MCP Overview

What is MCP?

The Model Context Protocol (MCP) is an open protocol that allows AI agents to interact with external tools and services. CAVOS Agents implements MCP client support, enabling agents to use tools from any MCP-compatible server.

How It Works

1. Design time — In the flow builder, add MCP servers to an agent node and select tools
2. Runtime — When the agent needs a tool, it calls the MCP server via streamable HTTP transport
3. Response — The tool result is fed back to the agent for further processing

Transport

CAVOS Agents uses the Streamable HTTP transport (MCP protocol v2025-03-26) for all MCP communication.

Architecture

Agent Node (flow runtime)
    ↓ tool call
MCP Client (streamable-http)
    ↓ HTTP request
MCP Server (external service)
    ↓ tool result
Agent Node (continues processing)

Built-in Servers

CAVOS Agents includes two built-in MCP servers:

Server	Port	Purpose
Memory	8001	Persistent knowledge graph storage
Sequential Thinking	8002	Structured reasoning and problem decomposition

These are proxied through an nginx reverse proxy for reliability.

Security

• SSRF protection — Built into the MCP client to prevent server-side request forgery
• Credential isolation — Secrets stored in Infisical (primary) or database (fallback), never exposed in flow definitions
• Timeout enforcement — Per-tool-call timeouts prevent hanging connections