What is MCP?

Why MCP matters for AI agents

Without MCP, each AI client needs one-off integrations for every system it wants to use. MCP gives the client a common interface for discovering tools and calling them. That makes it useful for codebase search, database access, ticket systems, email, calendars, internal APIs, and workflow automation.

• An MCP client is the AI application or runtime that wants to use tools.
• An MCP server exposes tools, resources, prompts, or workflows to that client.
• The model sees tool names and descriptions, then calls the right MCP tool when the task requires it.
• A managed AI agent runtime keeps those MCP connections available without every teammate configuring them locally.

MCP transports: Streamable HTTP, SSE, and stdio

MCP messages use JSON-RPC, but the messages can travel over different transports. For team and hosted deployments, the important remote transports are Streamable HTTP and SSE. Kognita supports both for custom MCP servers because some vendors and internal tools still expose SSE while newer remote MCP servers usually prefer Streamable HTTP.

How Kognita uses MCP

Kognita provides a managed AI agent runtime that is always on for the project. The runtime is provisioned once, refreshed when configuration changes, and connected to the project's codebase context and approved MCP tools.

• Kognita MCP: each project exposes an MCP endpoint so Claude Code, Cursor, Codex, Windsurf, and other MCP clients can query semantic codebase context.
• Built-in runtime MCPs: the managed project agent can connect to Kognita semantic search, Jira, database MCPs, and email MCP when enabled for the environment.
• Custom MCPs: you can add a public HTTPS MCP endpoint using Streamable HTTP or SSE, store headers encrypted, and provide guidance that tells the agent when to use it.
• Runtime refresh: after adding or editing a custom MCP, refresh the environment so the always-on agent discovers the new MCP tool group.

Current custom MCP support

Example: email MCP in the managed runtime

Suppose your team hosts an MCP server that sends approved transactional emails. Add it to Kognita as a custom MCP, then refresh the runtime. The managed agent can keep using Kognita codebase context while also calling the email MCP when a user asks it to send an approved message.

Name: team_email
Transport: Streamable HTTP
URL: https://mcp.example.com/mcp
Headers:
  Authorization: Bearer <email-mcp-token>
Agent guidance:
  Use this MCP only when the user explicitly asks to send an email.
  Confirm the recipient, subject, and body before calling the email tool.
  Do not use it for marketing blasts or unapproved bulk sending.

In the runtime, Kognita registers that server as a custom MCP tool group derived from the server name. For this example, the agent guidance would refer to a group like custom_team_email.

Connect external AI tools to Kognita MCP

This is the other direction of MCP. Instead of Kognita calling your custom MCP, your local AI tool calls Kognita's project MCP endpoint for codebase context.

Claude Code

claude mcp add-json kognita-mcp '{"type":"http","url":"https://mcp.kognita.co/projects/<project-uuid>/mcp","headers":{"Authorization":"Bearer <your-token>"}}'

Cursor

{
  "mcpServers": {
    "kognita": {
      "type": "http",
      "url": "https://mcp.kognita.co/projects/<project-uuid>/mcp",
      "headers": {
        "Authorization": "Bearer <your-token>"
      }
    }
  }
}

Codex

mcp_servers:
  kognita:
    type: http
    url: https://mcp.kognita.co/projects/<project-uuid>/mcp
    headers:
      Authorization: Bearer <your-token>

Next steps