MCP servers
Connect a Model Context Protocol server, then attach its tools and skills to the models that should use them.
What an MCP server provides
A Model Context Protocol (MCP) server publishes capabilities over a standard protocol. Inferada can connect to the server and make supported capabilities available during a chat completion.
Enable a server for the organisation, then attach it to individual models. Callers do not need to add tool definitions to each request. The gateway discovers the configured capabilities, executes tool calls, and returns results to the model.
Tools and skills
Inferada exposes MCP tools directly and treats compatible text resources as preloaded skills. Choose based on whether the model should call a capability or receive its content as context.
Functions the model can invoke when needed. The gateway sends the call to the MCP server and returns the result to the model within the same completion.
Reference material that the gateway loads into the conversation as context before the model responds.
Enabling a server
From Organisation → MCP servers:
- Click Enable on the server you want to attach.
- Enter credentials when required. The server definition determines whether it uses no authentication, a bearer token, or OAuth.
- Hit Test connection to confirm the gateway can reach the server and see its capabilities.
Next, open the model under Models and add the server to its MCP server list. Completions for that model can then use the server's capabilities.
Auth modes
The server definition selects one of these authentication modes:
Public server. Nothing to configure.
Paste a long-lived token. The gateway sends it on every request to the server. Rotate by entering a new value on the server's detail page; the previous one is overwritten.
Click Connect on the server's detail page. The gateway runs the standard authorisation flow with the upstream, stores the refresh token encrypted, and keeps the connection live as it rotates access tokens behind the scenes. Reauthorise the connection from the same page when needed.
Attaching to specific models
Enabling a server makes it available to the organisation but does not attach it to every model. Add it only to the models that need its capabilities.
Limits and safeguards
- All outbound calls to MCP servers route through the gateway's SSRF guard, which refuses private, loopback, and cloud-metadata addresses.
- Credentials are encrypted at rest and never returned in API responses. Rotate them by entering a new value; the previous one is overwritten.
- Tool-call chains follow the same per-model step cap as catalogue tools, so a misbehaving server can't spin the model in an infinite loop.
- The gateway caps preloaded skill content at 256 KB per request.
- If capability discovery fails, the gateway omits that server's capabilities for the request and continues the completion.