DeepSeek Harness Custom Provider: Gateway Setup
A custom model API needs a permanent provider ID, a documented endpoint URL, one supported protocol, a usable credential and at least one model ID. Use it for gateways absent from the installed catalog. Discovery can help populate the form, but you still need to save it and verify inference independently.
Sharing includes the public link only, without local progress or private configuration.
Write down the gateway contract
Use the gateway’s own documentation to identify its base URL, wire protocol and exact served model ID. OpenAI-compatible is not a guarantee of Responses support. The reviewed form offers Chat Completions, Responses and Anthropic Messages; create a separate provider for each protocol you use.
Create a permanent provider identity
Choose a lowercase provider ID before saving. Requests, sessions, defaults and credential references depend on it. Changing the label is safe; changing the identity requires creating a new provider and removing the old one. Do not reuse an ID belonging to another adapter.
Discover models or enter their IDs
Fetch available models uses the custom form’s current URL, protocol and key. Select only the needed entries, add them and save the provider. A built-in provider returns its installed catalog even if its URL was changed to a gateway. An endpoint with no supported listing format can still work with manually entered documented model IDs.
Verify protocol and compatibility separately
Test the saved model in a new session. If the key and model are correct but request fields are rejected, inspect the gateway’s accepted fields. The official guide describes compat.supportsDeveloperRole and compat.maxTokensField for gateways that require different request shapes; these are claims about endpoint behavior, not universal fixes. Preserve other provider entries when editing the whole config row.
First-party sources
Open the pinned or retrieved source before acting on a changed release candidate.