Connect Claude Code, Claude Desktop, or Cursor over MCP

MCP (Model Context Protocol) is how an AI agent learns what tools it can call. Letterstory's MCP server exposes every platform operation as a tool, so once your agent is connected you talk to it in plain English: "list my kernels", "research this topic and write it with Hazel", "onboard a phantom for acme.com". The agent picks the tools and does the polling.

You need an API key (KB-401), or a client that supports MCP sign-in (OAuth): leave the key header off and the client sends you to your browser to approve it. The server is:

https://app.letterstory.com/api/mcp

Claude Code

One command in your terminal:

claude mcp add --transport http letterstory https://app.letterstory.com/api/mcp \
    --header "x-integrations-key: ls_your_key"

Then start a Claude Code session and ask: "List my Letterstory writing kernels." You'll see it call the letterstory server. Run claude mcp list any time to confirm it shows Connected.

Claude Desktop

Open Settings → Developer → Edit Config and add the server to the mcpServers block:

{
	"mcpServers": {
		"letterstory": {
			"type": "http",
			"url": "https://app.letterstory.com/api/mcp",
			"headers": { "x-integrations-key": "ls_your_key" }
		}
	}
}

Restart Claude Desktop. The tools appear under the tools icon in a new chat.

Cursor

Cursor reads the same mcpServers shape. Open Cursor Settings → MCP → Add new server and paste the block above, or put it in .cursor/mcp.json in your project.

Any other MCP client

Use the same three facts: transport HTTP (streamable), URL https://app.letterstory.com/api/mcp, header x-integrations-key. If you have the CLI installed, letterstory mcp prints all of this filled in for you.

What the agent can see

Everything: tools/list returns the full catalog, and each tool carries a description and a typed argument schema, so a capable agent reads them and works out the sequence on its own. Your key's capabilities still apply, a tool the key isn't allowed to call returns an error the agent will report back to you.

First things to try

• "List my collections and tell me which ones have Auto-Pen turned on."
• "Give me ten topics I could write about, grounded in my positioning."
• "Research 'X' into a new draft in the Field notes collection, then have my default kernel write it and tell me when it's done."

More worked examples in KB-501 and a cheat sheet in KB-505.

Setting up a new account with your agent

If the account is empty, don't start from the prompts above. /docs/onboard (the Self-serve setup link on /docs) carries a copy-paste prompt for your agent. Rather than a fixed script, it has the agent ask the account what it needs (get_onboarding_plan), run the step that is ready (context, CMS, collection, planner, Search Console, phantom), and ask again until nothing is left. Create the key with the Onboarding + publishing agent preset so every step has the capabilities it needs. Paste the prompt into the same connected session and answer its questions.