Skip to main content
Choose between an MCP connection, an embedded Mind engine and the KaleidoAgent service. Skills customize behavior within these integrations. This page covers what to have in place first, which surface fits, and how each one gates the calls that move value.

​
Before You Start

If you have not installed anything yet, start with Installation. For an end-to-end path from an empty machine to a local model moving RGB assets on signet, follow Build a Local RGB Agent.

​
Pick a Surface

Whichever surface you pick, the first run follows the same order: a read-only call, then a wallet call, then a quote, and only then an execution. That order is diagnostic as much as it is cautious. Inspect the tool error at each step: market data can fail upstream, wallet reads need their own prerequisites, and a successful quote does not verify settlement capacity. Each surface carries its own walkthrough: MCP Servers and KaleidoAgent. KaleidoMind ships inside a host rather than running on its own, so its first run is the Desktop App installation. Skills change how an agent behaves on any of these surfaces, they are not a surface of their own.

​
What Confirmation Looks Like

The two agent surfaces gate spending differently, and it is worth knowing which one you are relying on.
A generic MCP host is the least protected path. If your client auto-approves tool calls, an LLM can spend from the configured wallet without asking you. Use a test network seed until you know how your client handles approvals.