Local and hosted describe where inference happens. They do not, by themselves, tell you how the surrounding application logs requests, which tools it can call, or whether its model is good at your task. Evaluate the complete path from editor to endpoint.
Inspect the shape of your setup
You operate the inference environment
Check hardware requirements, model support, endpoint compatibility, and the time needed to keep the service working. Confirm whether the surrounding application or its tools still contact external services.
A provider operates inference
Check model access, authentication, request limits, retention terms, and failure behavior. Verify the current provider settings that apply to your account before sending repository content.
Different tasks take different paths
Define a routing rule before use: which material may reach which endpoint, and who chooses. Keep the selected endpoint visible so a convenient fallback does not become an accidental data transfer.
Try an ordinary task
Use a small, non-sensitive project to check connection behavior and response quality. Observe both a successful request and a connection failure. Confirm that the application identifies the chosen model and does not silently switch to a different endpoint.
Record the actual boundary
Write down what leaves the machine, where credentials live, and which tools can access the network. Local inference is not a privacy audit; hosted inference is not a single uniform policy. Your chosen application and provider configuration determine the details.
