Turn any CognitionDesk assistant into a custom MCP server (with RAG) in 2 minutes

If you use Claude Code, Claude Desktop or GitHub Copilot, you already live in the MCP ecosystem. Here's a trick most people don't know: every CognitionDesk assistant is also an MCP server. That means you can build an agent with RAG over your own documents in a web panel — zero code — and two minutes later your coding agent is calling it as a tool.

Why you'd want this

• Project memory for your coding agent: upload your architecture docs, API specs and runbooks to an assistant; your Claude Code session can now query them semantically instead of you pasting context.
• Domain experts on demand: a "billing-rules expert" or "compliance expert" assistant, maintained by whoever owns that knowledge, callable from any dev's editor.
• No infra: no server to run, no vector DB to manage. The MCP endpoint is hosted.

1. Create the assistant

At app.cognitiondesk.com, create an assistant and upload the documents you want it to know. Add tools if you like (API connections, inventory, appointments — anything the panel offers). Generate an API key with platform access.

2. Add it to your MCP config

Every assistant lives at /mcp/:assistantId (JSON-RPC over HTTP, authenticated with x-api-key). Add it to your project's .mcp.json:

{
  "mcpServers": {
    "my-knowledge-agent": {
      "type": "http",
      "url": "https://api.cognitiondesk.com/mcp/YOUR_ASSISTANT_ID",
      "headers": { "x-api-key": "YOUR_API_KEY" }
    }
  }
}

Restart your session and the assistant appears as an MCP server: ask it questions, and it answers with retrieval over the documents you uploaded — citations included.

3. Iterate without redeploying

This is the part that compounds: the MCP config never changes, but the agent behind it keeps getting better. Upload new docs, tune the system prompt, add tools, switch the model from GPT to Claude to a local Ollama model — the panel is the control plane, and every MCP client picks up the changes instantly.

The bigger picture

MCP turned AI agents into an ecosystem of interoperable tools. CognitionDesk's take: you shouldn't have to write a server to participate. Describe your agent, feed it knowledge, and it's simultaneously a web widget, a Telegram bot, an API — and an MCP server for every other agent you use.

Build one, plug it in, and tell us what you connected it to.