Build with AI
The three ways artificial intelligence touches a Xuda project, from the assistant inside Studio to an external agent over MCP.
Xuda puts artificial intelligence to work in three different places, and it helps to know which one a task calls for before you start.
The AI App Builder, in Studio
Every Studio project opens with a chat panel next to the inspector: describe what you want, and it scopes the change to the whole project or to whatever is selected, then applies it. It is the fastest way to scaffold a component, explain an unfamiliar project, or wire a workflow without leaving the editor.
See Prompting patterns for what to say to it.
Agents you build, not just chat with
Studio also has its own prog type for this: ai_agent. Instead of hand-assembling a prompt and a
model call, you describe a goal in a couple of sentences, and Studio proposes a plan, asking
clarifying questions first when the goal leaves real gaps (who uses it, what sources it should
draw on, what it should produce). The result is a standalone agent that can be published into the
account's AI Agentic Workspace for anyone with access to chat with, the same area covered by
AI Workspace on the API side.
A published agent is not only something a person chats with. Another agent's own tool list can carry an AI Agent tool that reaches one specific published agent, or every agent the project has published (never itself, and a utility agent may not reach an assistant): the model can then call that agent as a tool, delegating the request to it, the same way it calls a table or an installed plugin.
MCP, for anything outside the project
Point an external, MCP-aware client (an IDE agent, a chatbot, a script) at
https://xuda.ai/mcp with an API key, and every CPI method that key can call becomes a typed tool
for it, no separate integration. The same connection also hands the agent the guide covered on
The Studio LLM guide: the prog document model, the
xu-* attributes, the workflow actions, and the landmines that trip up a first attempt, merged
from the deployed platform's own rules so it never drifts from what actually validates.