Langflow: OpenAI API Base, typed model names, and the server-key endpoint guard

[apigoto]

I work on the APIGOTO team. APIGOTO is an independent OpenAI-compatible gateway, not affiliated with OpenAI, Anthropic or Google. This applies to any OpenAI-compatible endpoint.

1. Confirm the endpoint and key

curl https://api.china-aiapi.com/v1/models -H "Authorization: Bearer $KEY"

2. OpenAI component (OpenAI bundle)

1. Add the OpenAI component to the flow.
2. OpenAI API Key: paste the key, or a global variable holding it (but see the guard below).
3. Show the advanced fields and set OpenAI API Base to https://api.china-aiapi.com/v1. Stop at /v1; the SDK adds /chat/completions.
4. Model Name is a combobox: type MODEL_ID even if it isn't listed.
5. Run it in the Playground.

Two things from the component source:

• It doesn't read OPENAI_API_BASE / OPENAI_BASE_URL. An empty field means https://api.openai.com/v1, so fill it on every component.
• Server-key guard. If the key resolves to a value from the server's environment (e.g. a variable seeded from env, or the SDK picking up OPENAI_API_KEY), Langflow won't send it to a non-default host. You'll see "Refusing to send a server-provisioned API credential to the non-default endpoint…". Fix: give the component an explicit key, or have the operator add the host to LANGFLOW_PROVIDER_CREDENTIAL_ALLOWED_HOSTS. A custom endpoint also has to be https unless allowlisted.

3. Alternative on 1.12: OpenAI Compatible provider

Profile icon → Settings → Model Providers → OpenAI Compatible. Enter Base URL and API Key (or set OPENAI_COMPATIBLE_BASE_URL / OPENAI_COMPATIBLE_API_KEY), then Save. Models are discovered from /v1/models; enable them and pick one in any Language Model or Agent component.

Troubleshooting

• 404s: /v1 missing, or doubled by pasting a full endpoint.
• Model not in the provider list: the endpoint's /models doesn't return it. Use the OpenAI component and type the ID.
• Agent ignores tools or fails mid-run: tool calling and streamed tool-call chunks differ between model families. Try another MODEL_ID before debugging the flow.
• JSON Mode errors: not every model behind a compatible endpoint supports response_format.

We use one key for several model families, with upstream failover (at most 3 retries) and a per-request log. Pay-as-you-go, BYOK supported: www.china-aiapi.com

For shared deployments, do you prefer the provider or per-component base URLs?