One endpoint,
every model.
Tare is an OpenAI-compatible gateway in front of every model you use — ours and the ones you bring yourself. One base URL, one key, and every token accounted for.
Channels
A channel is one upstream you brought yourself: a base URL plus a credential.
Console → Channels → Add channel.
Base URL and dialect
Give the URL up to /v1, for example https://api.example.com/v1. The endpoint's dialect decides
what gets appended:
| Dialect | Chat endpoint | Model list | Credential |
|---|---|---|---|
| OpenAI-compatible | {base}/chat/completions | {base}/models | required |
| Anthropic | {base}/messages | {base}/models | required |
| LM Studio (local) | {base}/chat/completions | {base}/models | not used |
| Custom | nothing appended — give the full URL | give the full URL | required |
Custom covers upstreams that do not follow that path convention: Azure requires a mandatory
api-version query string, Gemini puts the model name in the path. Without that option those
upstreams fall into the OpenAI row and return a 404 with no obvious cause.
[!WARNING] LM Studio runs on your machine, and the connection is opened from Tare’s servers, not from your browser, so
localhostis not reachable. Enable "Serve on local network" and supply an address reachable from the public internet.
The dialect rewrites the request body
Choosing Anthropic changes the request body, not just the URL: Tare translates OpenAI shape to
Anthropic shape outbound and back inbound. System messages lift out of messages, tool calls
become content blocks, images change shape, and max_tokens — optional for OpenAI, mandatory
for Anthropic — is filled in when you did not send one.
Everything else passes through untouched; translation happens only where passing through is impossible.
Model list
Press Browse upstream models. Tare asks the upstream what it offers, with prices where the upstream publishes them, and follows the upstream's paging to the end — Anthropic returns 20 per page and Gemini 50, so a single page would silently be a fraction of the catalogue.
Very long catalogues (aggregators with thousands of models) are cut off, and when that happens the dialog says so. Search only reaches what is listed, so type the exact name if you do not find it.
[!NOTE] This is a suggestion, not an allow-list. Private deployments, preview builds and aliases never appear in a provider's list but remain usable; entering an unlisted model name is always permitted.