Connecting an AI client over MCP
Add Xuda as a remote MCP server to Claude, Cursor, or any other MCP-aware client.
This page is about the client side: getting Claude, Cursor, or anything else that speaks MCP, talking to your Xuda account. For the protocol itself, tool naming, session lifecycle, error shape, see MCP.
What you need
One thing: an API key. Create one from the account's Developers area (full access reaches
everything the key's owner can, custom access reaches only an allowlisted set of methods), see
Authentication. There is no separate MCP credential, the same key
works over plain REST at /cpi/<method> and over MCP.
Every account API key, and every Full Stack VPS box's own auto-issued key, point at the same
address: https://xuda.ai/mcp. A Full Stack VPS key additionally exposes that box's installed
plugin tools alongside the ecosystem methods, see Full Stack VPS.
Connect a client
Cursor reads its snippet from .cursor/mcp.json (project) or its global MCP settings. Claude
Desktop and claude.ai add a remote server from Settings, under Connectors: a name, the URL, and
the same Authorization header. Any other MCP client needs only the same two facts: a streamable
HTTP endpoint at https://xuda.ai/mcp, and the key on every request as
Authorization: Bearer YOUR_API_KEY.
Every account's Developers, API keys page has a Connect card under the key list. Its format
picker switches one block between .env, Next.js, MCP, cURL, Node.js and PHP, each already
filled in with the endpoint and the first key in the list, which the block's header names. The
.env format carries XUDA_MCP_URL next to the REST base and the key, and the MCP format is the
Claude Code command above. The key shows masked until you reveal it; Copy always takes the real
value.
What the agent gets on connect
The first request on a new connection is an MCP initialize call. Xuda's response carries an
instructions field built from the Studio LLM guide, the
same prog document model, xu-* attribute reference and workflow rules covered on that page,
merged from the platform's own live registries so it can never describe a rule the deployed
validator does not actually enforce. Most MCP clients forward instructions straight into the
model's context, so the agent already knows how a Studio prog is put together before its first
tool call, no separate briefing required.
Two tools are worth calling directly rather than relying on what came in at connect time:
get_studio_guide rereads any topic on demand (components, tables, programs, javascript,
attributes, anatomy, gotchas), and get_prog_examples returns working, curated example
progs to pattern a build on.
A dedicated endpoint for one API program
The methods above are the whole account's catalog behind one key. When you want an external system
to reach exactly one endpoint of one project instead, deploy it: Deploy > Deployments > Server
lists MCP Server alongside Cron, Api, Node and Emitter. Its Launch Settings step picks one of
the project's own api programs as the target, then the usual datacenter, performance tier and
naming steps ship it as its own endpoint. See Your first deploy
for the shared parts of that wizard, and API for what an api
program is.
mcp is listed among the platform's beta deploy types today, alongside router, android and
emitter.