Connect Kapten to Claude Code, Claude Desktop and Cursor

What you are connecting

Kapten exposes its platform through the Model Context Protocol (MCP), an open standard for giving an AI assistant a structured set of tools. Once connected, your assistant can read the state of your environments, ask the Kapten crew for plans, and, with your approval, apply them.

The assistant stays the same one you already use. Kapten adds the context it is missing: what is running, where, in which version, and what it costs. Read the MCP overview for the bigger picture.

Early access note. The endpoint and tool names in this guide are illustrative while Kapten is in early access. The exact endpoint, supported clients and available actions for your account are confirmed during onboarding.

Before you start

You will need:

• A Kapten early-access account. Request one here if you do not have it yet.
• At least one project with a connected cloud account, or a repository you want to deploy.
• A client that supports remote MCP servers over HTTP. The examples below cover Claude Code, Claude Desktop, Cursor and VS Code; any compliant client follows the same pattern.

You do not need to create an API key and paste it into a configuration file. Authentication happens in the browser the first time a tool is called.

Claude Code

Add the server from your terminal:

claude mcp add --transport http kapten https://mcp.kapten.io/mcp

Start a session with claude and run /mcp to see the server and its status. The first time Claude calls a Kapten tool, it opens a browser window to sign in.

To make the server available to everyone working in a repository, add it with project scope so the configuration is checked in alongside the code. Keep it user-scoped if the repository is public.

Claude Desktop

Open the configuration file from Settings → Developer → Edit config and add:

{
  "mcpServers": {
    "kapten": {
      "type": "http",
      "url": "https://mcp.kapten.io/mcp"
    }
  }
}

Restart Claude Desktop. Kapten appears in the tools menu of a new conversation.

Cursor

Create or edit .cursor/mcp.json in your project, or the global file in your home directory:

{
  "mcpServers": {
    "kapten": { "url": "https://mcp.kapten.io/mcp" }
  }
}

Open Settings → MCP to check that the server is enabled.

VS Code

Create .vscode/mcp.json in your workspace:

{
  "servers": {
    "kapten": { "type": "http", "url": "https://mcp.kapten.io/mcp" }
  }
}

Start the server from the file’s inline action, then use it from the chat in agent mode.

Sign in and choose a scope

On the first tool call, Kapten opens a sign-in page. After signing in you choose what this connection can see and do:

• Organisation and projects. Limit the connection to the projects you are working on.
• Environments. Pick which environments are in scope. A common starting point is staging only.
• Tier. The highest tier this connection may use: Observe, Plan or Act.

The scope belongs to the connection, not to the conversation. Asking the assistant nicely does not widen it; changing it requires signing in again. You can review and revoke connections from your Kapten account at any time.

The three tiers

Every Kapten tool belongs to one of three tiers. Knowing which one a request falls into tells you what will happen.

Start with Observe and Plan. Move a connection to Act once you have read a few plans and are comfortable with what they contain. The guardrails guide explains the controls in detail.

Five first prompts

These are written to be safe on a fresh connection: the first three only read, the last two only plan.

1. Get your bearings.

List my Kapten projects and environments. For each environment,
show the services, their current version and who owns it.

2. Understand the last change.

What changed in the last deployment to staging? Show the build,
the image, the config diff and whether health checks passed.

3. Investigate a service.

The API in staging returned 5xx errors in the last hour.
Read its logs and metrics and tell me what you think happened.
Don't change anything.

4. Plan a new environment from this repository.

Analyse this repository and plan a staging environment for it on
our connected AWS account. Show me the services you found, the plan
and the monthly estimate.

5. Start a workflow.

Run a production-readiness review on the staging environment
and give me the findings sorted by severity.

The fourth and fifth prompts start multi-agent workflows. You will see Kapten hand parts of the request to specialists, then return a single plan or report.

What the tool calls look like

Your assistant shows each Kapten tool call before it runs, with its arguments. Names follow the tier they belong to: get_environment, tail_logs and get_costs read; plan_environment, plan_release and review_security plan; apply_plan, deploy and rollback act. The full illustrative list is on the MCP page.

Most clients let you allow read-only tools automatically and confirm the others each time. That is a sensible default: allow Observe tools, review Plan calls, and always confirm Act calls yourself.

Troubleshooting

• The server shows as disconnected. Check that the URL is exactly the one given during onboarding and that your client supports HTTP transport.
• The assistant says it has no Kapten tools. Restart the client after editing its configuration, and check that the server is enabled.
• A tool returns “out of scope”. The connection does not include that project, environment or tier. Sign in again and widen the scope deliberately.
• Sign-in loops. Clear the stored credentials for the server in your client and connect again.

Next steps

• Read How multi-agent workflows run to understand what happens after you ask for a plan.
• Browse the workflow library for prompts that cover deployment, incidents, security and cost.
• Set your guardrails before moving a connection to Act.