Adapters normalise each provider to one canonical request and response. Your credentials are resolved at the moment a provider is called — never carried through the planner, never written to a manifest, never logged.
| Provider | Status | Notes |
|---|---|---|
| Anthropic | available | Chat, streaming, tool calls. Real list pricing with recorded provenance. |
| OpenAI | adapter ready | Adapter implemented; ships with no pricing rows, so endpoints fail closed under a cost ceiling until priced. |
| Self-hosted / OpenAI-compatible | supported | Any endpoint speaking the same shape drops in via a base URL — required where a data class must not leave your estate. |
| Mock | testing | Deterministic adapter for CI. Excluded from quality measurement, because a mock's grades describe the harness, not a model. |
Sampling parameters differ by model — some current models reject a temperature outright. Capability flags drive what is sent, defaulting to omit when unknown: omitting an optional parameter degrades a request, sending an unsupported one fails it.
Rate-limited means back off and retry. Out of credit means retrying cannot possibly work — and marking it retryable sends the fallback chain through every remaining endpoint to reproduce the same failure. We distinguish them.
A thinking model that spends its whole budget reasoning returns no text. That is a mis-provisioned request, not evidence about the model, so it is excluded from quality measurement rather than graded as a failure.
No markup, because Planverity never holds your provider balance.
Get started