A C4 MCP server gives an AI client tools to inspect and work with architecture. Its value depends on the model behind it: stable element identity, named relationships and clear boundaries make retrieved context useful. A collection of screenshots does not supply the same structure.
Start with the model the agent will read
Inspect this public e-commerce model as a human first. Identify the payment boundary and the services involved in creating an order. Use that question to test an agent’s answer against the actual model rather than accepting a fluent explanation.
The public viewer does not require authentication. MCP access to your organization is a separate authenticated workflow; opening this embed does not grant access to its source workspace.
UXXU connection requirements
The endpoint pattern is:
https://app.uxxu.io/api/mcp/<organizationId>
Authorization: ApiKey <key>
Use a client that supports the configured Streamable HTTP connection and custom authentication headers. Put the key in the client’s secret settings, not in a committed example file. UXXU’s current endpoint uses API keys; an OAuth-only connector cannot use that setup directly.
Follow the MCP setup guide for client-specific configuration. The organization in the URL and the key’s access determine the scope of the tools. A connection example alone is not an end-to-end compatibility test for every client version.
Verify reads before writes
Ask the client to discover its tools, list the permitted projects and read a selected model. Ask it to return the IDs and descriptions of the relevant elements, then explain a dependency with evidence from the returned relationships.
Next, request a proposed change: rename a provider or add a documented integration. Confirm the target workspace and inspect the proposal before applying it. After a write, read the model again and check the related views. If a request times out, inspect the result before retrying so you do not create duplicate elements.
Use a repeatable exercise
Download the checkout specification. Build a small model from those responsibilities in a workspace you control. Have the agent identify which component talks to the Payment provider, then propose a provider rename and show the affected views.
Use the downloadable architecture review prompt to capture actual model IDs, repository evidence and unresolved questions in the same response. A successful read exercise returns verifiable identities and relationships; a plausible explanation without those references is not a completed check.
The C4 model tool provides the visual side of that workflow. For a larger example of keeping an agent grounded in boundaries and evidence, see C4 architecture for AI coding agents.