Each MCP server exposes one capability and has its own OAuth scope and Neuron privileges. Connect only the server your AI application needs.
Enabled servers and tool counts differ by deployment. Use the MCP discovery response from your Neuron instead of assuming every item below is available.
Select the smallest server
Connect an agent only to the server it needs. There is no joined /MCP endpoint.
Neuron declares MCP safety hints on each tool:
- whether it can modify the environment;
- whether it can destroy data;
- whether repeated calls are idempotent;
- whether it accesses the open world.
These annotations help a client plan tool calls; they do not grant access. Authentication and Neuron privileges determine what the account can do.
Resource isolation
File and XMPP resources are account-specific and require authentication. A resource URI obtained by one agent must not be treated as a shareable public URL.
Extension model
Packages can register additional HttpMcpServerResource implementations. An extension should publish its own OAuth resource name, scope root, narrow privilege leaves, and accurate safety annotations. See Build an MCP server. 
