-
Notifications
You must be signed in to change notification settings - Fork 1
BYOK Guide
Bring Your Own Key (BYOK) lets a caller select an AI provider for an active Trinity tool without changing the hosted server's default routing.
Provider catalogues and model IDs expire. Before configuring BYOK, check:
-
/healthfor hosted per-stage routing, catalogue freshness, construction fallback, and runtime-failover state; -
/setupfor current provider configuration; - the server card for live discovery;
- Current Production Status for the deployed release boundary.
Do not choose a model from an old Wiki table.
The current remote BYOK catalogue has six providers:
| Provider | Provider ID |
|---|---|
| Google Gemini | gemini |
| Anthropic | anthropic |
| OpenAI | openai |
| Groq | groq |
| Cerebras | cerebras |
| Mistral | mistral |
ollama is a separate local/self-hosted provider. It works only when the server
process can reach the caller-managed Ollama service; the public hosted service
cannot use an Ollama process running on your laptop.
xAI is not part of the current provider catalogue.
Use the model catalogue reported by the live service. The server applies a freshness contract to the six remote catalogues, but provider-side availability can still change between verifications.
With no BYOK parameters, the hosted service selects its configured provider for each X, Z, and CS stage. The three stages are not guaranteed to use three different providers or model families.
Pass llm_provider and api_key to an active Trinity tool. For
run_full_trinity, the global pair applies to all stages unless a per-stage
override is present.
Example request in natural language:
Use
run_full_trinitywith my Groq provider credentials. Keepsave_to_history=falseand report the provider/model and quality marker for every stage.
run_full_trinity also accepts:
-
x_provider+x_api_key -
z_provider+z_api_key -
cs_provider+cs_api_key
Per-stage values take precedence over the global pair. This lets a caller deliberately select different providers, but diversity should be verified from the returned stage metadata rather than assumed from configuration intent.
- Keys are request-scoped and are not intentionally stored as account data.
- Provider failures are logged without raw keys or raw SDK response bodies.
- The selected provider receives the prompt, optional context, and any prior stage output needed for that inference.
- Provider processing remains governed by that provider's terms, region, retention settings, and the caller's account agreement.
- Never put a key in a prompt, tracked file, issue, discussion, screenshot, or shared transcript.
- Prefer an explicit provider ID. Key-prefix detection is convenience behavior, not a durable integration contract.
Read the live Privacy Policy before sending personal, confidential, regulated, or third-party data.
Keep these mechanisms separate:
- Construction fallback chooses a provider only when the configured hosted provider cannot be constructed.
- Runtime failover would move an in-flight stage between providers only when the separately gated live switch is enabled.
- Trinity degradation preserves completed trustworthy stages and marks a failed stage unavailable.
The current runtime-failover state is reported by
/health. Never describe a configured
fallback chain as active runtime failover when that switch is false.
A BYOK error must not be interpreted as successful real inference. Check the returned stage quality, provider/model attribution, normalized failure class, retryable flag, and recovery guidance. An incomplete run remains incomplete even when other stages succeed.
Before depending on a BYOK route:
- Confirm the provider and model are present in the live catalogue.
- Run a benign, non-sensitive standalone test.
- Verify returned provider/model and
realquality. - Test one red-line or failure case appropriate to your use.
- Confirm
save_to_history=falseunless aggregate history is explicitly intended. - Re-check after a provider retires or renames a model.
VerifiMind PEAS documentation · Current status · Live health · Public statements · MIT License
Runtime versions, models, routing, tool availability, policies, metrics, and deployment facts are owned by their linked live or release-bound sources.
Start Here
Operator Playbook
Textbook
Trust & Transparency
Evidence & Research
Project Links