mcp

MCP overview

The Roxels MCP package lets an AI agent — Claude Code, Cursor, or any MCP-compatible client — set up, inspect, and improve your Roxels templates.

If you've ever stared at a "create template" UI and wished you could just say "I want a customer onboarding conversation that captures name, role, and use case, and fires a webhook to my CRM" — that's what MCP makes possible.

Who this is for

Developers setting up Roxels for the first time. The MCP assistant pulls in best practices and points out what your config is missing.
Template authors iterating quickly. Update goals, schemas, outputs, and skills through natural language.
Anyone debugging a template. "Why isn't this webhook firing?" is a fair question for the assistant; it can read the configuration and explain.

What MCP gives you

When you connect Claude Code (or another MCP client) to Roxels, the assistant gains a set of tools it can call on your behalf:

List, create, update, and inspect templates.
Start conversations (sandbox runs) to test a template.
Read past sessions — transcripts, captured data, where things went wrong.
Configure webhooks and verify they're wired correctly.
Generate the embed snippet for your site.
Assess template quality and recommend improvements based on real session data.

The single most useful MCP tool is guided_setup_questions. Calling it on any template returns a list of concrete, actionable recommendations — what's missing, what's misconfigured, what real session data suggests you should fix. It's the closest thing to a personal Roxels consultant in your editor.

How it works

You install the roxels-mcp package (a small Python CLI).
You configure your MCP client (Claude Code, Cursor, Claude Desktop, etc.) to point at it.
You authenticate it with your Roxels API key.
The assistant can now read and modify your templates as part of the conversation you're having with it.

See Install for the exact setup.

A typical session

A developer is setting up their first template:

You: I want a Roxels template for screening sales leads. It should capture company name, company size, what they're trying to solve, and budget. Send the result to my Hubspot via webhook.

Assistant: I'll create a template with four goals (company_name, company_size, problem, budget) and a webhook output. What's the URL for your Hubspot endpoint?

You: ...

Assistant: Done. Template tpl_xyz is created. I notice you didn't enable the persistent launcher — do you want this to live on a specific page, or should it be triggered manually? Also, the budget goal should probably be structured with a schema so Hubspot gets a clean number — want me to set that up?

This is the design point: the assistant knows the product, sees your config, and steers you toward what works.

What about the dashboard?

Both work. The dashboard is the right surface when:

You want to visually browse templates and sessions.
You're handing off to a non-technical teammate.
You want to make a quick edit without context-switching to your editor.

MCP is the right surface when:

You're authoring a new template from scratch.
You want batch operations ("update these 5 templates to use the new skill").
You're debugging and want the assistant to read transcripts alongside config.
You want recommendations grounded in your actual session data.

You can use both on the same templates.

What it isn't

It's not a separate Roxels account. It uses your normal API key and respects your org's permissions.
It's not a third-party service. It's published by Roxels.