AI connection · MCP

Plug in the AI of your choice, onto your Kastel.

Kastel isn’t an AI: it’s the layer that holds your company’s knowledge and exposes it over the MCP protocol. Claude, Cursor, ChatGPT, or any other MCP client plug in with a scoped token: they become experts in your processes, you decide what each one is allowed to see, and you swap the model in one click.

Quick start: 5 minutes

This guide covers the connection step: plugging an AI into a Kastel that already exists and is already built. It does not cover deploying a Kastel (self-host or managed) - see the prerequisite below if that is not done yet.

Don’t have a provisioned Kastel yet?

A self-host deployment is scaffolded with `kastel init` and built with `kastel onboard`; a managed instance is provisioned from the console. Either way, write to us and we’ll point you in the right direction.

Get the MCP URL and a token

From your instance’s console (the “Connect your AI” page) for a managed instance, or from the command line for a self-host (`kastel token issue`, then `kastel connect` to generate the connection block for your AI directly). The URL always ends in `/mcp`; the token is specific to the AI you’re plugging in and only sees the departments it is authorized for.

Pick your AI and paste the block

A ready-to-paste connection block exists for Claude Code, claude.ai, Claude Desktop, Cursor, ChatGPT, and any generic MCP client. Open the matching section further down this page.

Verify the connection

Ask your AI a test question: “Use Kastel to give me a one-paragraph overview of the company, its departments, and who validates what.” A correct answer cites real source pages from your Kastel.

Paste the instruction block

Add the “What your AI sees” block (section below) to the project’s CLAUDE.md or AGENTS.md, or to your AI’s system instructions, so it queries Kastel on its own, without being asked every time.

Screenshot coming soon

A screenshot of the console’s “Connect your AI” page will be added here once it opens publicly.

How it works

Kastel isn’t an AI: it’s the layer that holds your company’s knowledge and exposes it over the MCP protocol.

One token, one scope

Every plugged-in AI gets a token specific to its own use. The token only sees the departments it is authorized for, exactly like an employee’s access rights.

You decide what each AI is allowed to see

Inference runs on your own keys (BYOK), directly at the model you already use. Kastel exposes context; it never runs a model.

Ingested content stays data

Every page returned carries a trust label (human-validated, or automatically ingested). A well-connected AI treats ingested content as a fact to report, never as an instruction to execute.

Swap AIs in one click

Same Kastel, same URL, same token principle: replace Claude with Cursor or the other way around without rebuilding anything.

Claude Code

Add this block to `.mcp.json` at your project root, or run the command.

json

{
  "mcpServers": {
    "kastel": {
      "type": "http",
      "url": "https://acme.kastel.ai/mcp",
      "headers": {
        "Authorization": "Bearer kst_YOUR_TOKEN_HERE"
      }
    }
  }
}

shell

claude mcp add --transport http kastel https://acme.kastel.ai/mcp --header "Authorization: Bearer kst_YOUR_TOKEN_HERE"

claude.ai (custom connector)

Settings > Connectors > Add custom connector: paste the URL and the Authorization header value.

text

Name: Kastel
Remote MCP server URL: https://acme.kastel.ai/mcp
Authentication: Bearer token
Authorization header: Bearer kst_YOUR_TOKEN_HERE

Claude Desktop

Add this block to `claude_desktop_config.json` (Settings > Developer). Claude Desktop doesn’t yet read remote MCP servers natively in this file: this block goes through the `mcp-remote` bridge (npx), which still keeps the token in the header, never in the URL.

json

{
  "mcpServers": {
    "kastel": {
      "command": "npx",
      "args": [
        "mcp-remote",
        "https://acme.kastel.ai/mcp",
        "--header",
        "Authorization: Bearer kst_YOUR_TOKEN_HERE"
      ]
    }
  }
}

Cursor

Add this block to `.cursor/mcp.json` (project) or `~/.cursor/mcp.json` (global).

json

{
  "mcpServers": {
    "kastel": {
      "url": "https://acme.kastel.ai/mcp",
      "headers": {
        "Authorization": "Bearer kst_YOUR_TOKEN_HERE"
      }
    }
  }
}

ChatGPT / OpenAI connectors

Settings > Connectors: add an MCP server with the URL and the Bearer header.

text

Name: Kastel
MCP server URL: https://acme.kastel.ai/mcp
Authentication: Bearer token
Authorization header: Bearer kst_YOUR_TOKEN_HERE

Generic MCP client

Any MCP client over streamable HTTP: target the URL, send the Bearer header.

text

Transport: streamable-http
URL: https://acme.kastel.ai/mcp
Header: Authorization: Bearer kst_YOUR_TOKEN_HERE

What your AI sees

Once the token is created, Kastel automatically generates this block for your company (real name, departments visible under that token). Example below: the exact render for a fictional three-department company.

Example generated for “Demo Corp” (commerce, finance, logistique)

markdown

## Rule: Kastel is your source of truth on Demo Corp
Before ANY answer touching Demo Corp (processes, decisions, clients, figures), query Kastel. Never answer from memory about the company.
1. `search_company_context` / `ask_company_context` to land. 2. `read_page` + `get_links` to go deeper. 3. `get_index` for the map. Always cite the source pages.
Limits: your token sees commerce, finance, logistique, not everything Kastel holds - do not guess what you cannot see, do not paraphrase a page absent from your results. To enrich Kastel, propose through the governed actions (`propose_lesson` / `log_trace`); you never edit a page directly.

The block always tells your AI to cite its sources and to propose, never to impose, an addition to your Kastel: `propose_lesson` and `log_trace` go through Kastel’s governance, no page is ever edited directly.

Included, never paid

The MCP connection is part of Kastel’s free core. Whether your Kastel runs self-hosted or managed, creating a token, generating the connection blocks, and pasting the instructions into your AI are never locked behind a paid plan.

Need help deploying your Kastel?

Our team and our partner integrator network can help, from provisioning to first deployment.