Base URL override
By default Promptly sends requests to the provider's own endpoint — you never type it. The Base URL override field, under Advanced in the Add modal, points a credential somewhere else instead.
Leave it blank unless you have one of these:
- a corporate proxy or LLM gateway your requests must go through,
- a self-hosted or regional endpoint that speaks the same OpenAI-compatible API,
- a vendor-compatible service you want to reach using one of the five provider profiles.
What the override has to provide
Promptly appends its own paths to whatever you enter — /chat/completions for messages, and the provider's model-listing path for the picker. So the URL must be the API base, not a full endpoint: https://gateway.example.com/v1, not https://gateway.example.com/v1/chat/completions.
The target must speak the same request and response shape as the provider you selected, and accept your key the same way. A gateway that only forwards is fine; one that rewrites the payload usually isn't.
🖼️ [Image] — The Base URL override field with its hint: "Only needed for a proxy or self-hosted gateway."
Internal addresses are blocked
The override is a URL your workspace can make Promptly's servers call, so it's validated as untrusted input. Requests are rejected when the target is:
- not
httporhttps, or not an absolute URL, - loopback or wildcard —
localhost,127.0.0.1,::1,0.0.0.0, - a private range —
10.x.x.x,172.16–31.x.x,192.168.x.x,169.254.x.x, - a cloud metadata endpoint —
169.254.169.254,metadata.google.internal.
This is standard SSRF protection: without it, a saved base URL could be used to reach services on our internal network. The check runs when you save the credential, when the model list is fetched, and again at connection time against the address the hostname actually resolves to — so a public hostname pointing at a private IP is refused too, not just a literal one.
Practical consequence: you can't point Promptly at a model server running on your own laptop or inside your own VPC. It has to be reachable from the public internet.
Changing it later
Editing the base URL is treated like re-keying — the cached model list for that credential is dropped and the credential is re-checked, because a different endpoint may expose a different set of models and a different account entirely.
If the new endpoint is wrong or unreachable, you'll see it immediately: the model picker falls back to defaults with the connection error in brackets, and verification reports Provider down. See Key statuses.