Model providers

Bring your own intelligence means Achiral does not have to be the model vendor.

Achiral supplies the Achiral Account: memory, retrieval, permissions, routing, app runtime, and audit trail. Your team can supply the model endpoint that does the generation.

Model providers are the model accounts your team connects to Achiral. Chiro, Memory-Native Apps, memory extraction, and app narratives use those providers for generation.

Your team owns the provider account, the API key, the provider bill, and the agreement with that provider.

Routing order

Achiral tries the cheapest path that can answer the request:

1. Use memory or deterministic app projections when no model call is needed.
2. Use your connected model providers in the order an Owner or Admin set.
3. Use Achiral-hosted fallback only when fallback is allowed and your team's providers cannot serve the request.

Achiral does not choose its own models because they might give a better answer. Fallback is an availability path.

Chiro keeps each request, including its tool steps, on one provider. If that provider fails partway through a request, the reply stops and asks you to send the message again. Chiro does not switch models in the middle of a request.

Provider types

The settings page supports:

• Anthropic
• OpenAI
• Google Gemini
• OpenAI-compatible HTTPS endpoints that pass Achiral's endpoint checks

OpenAI-compatible support is the broad path. If your model service speaks the OpenAI chat completions shape over HTTPS, your Achiral Account can usually point Chiro at it after the endpoint passes safety checks.

Owners and Admins can add providers, verify them, reorder them, remove them, and turn Achiral fallback on or off.

Consumer subscriptions are not enough. Claude Pro, Team, or Enterprise and ChatGPT Plus, Team, or Enterprise do not give Achiral an API key and cannot serve as model providers.

What your provider sees

When a request needs generation, Achiral sends the provider only what that request needs:

• the user's request, or the app narrative request;
• relevant excerpts from team memory and connected sources;
• recent conversation or app-run context; and
• the instructions and output format.

Your provider agreement controls how that provider handles the request.

Achiral does not train or fine-tune models on your team's memory, app specifications, app outputs, prompts, or provider responses. See the Data Processing Policy.

What stays on Achiral

Achiral still runs the Achiral Account.

That includes connected-source ingestion, memory storage, retrieval, permissions, app specifications, deterministic projections, run records, audit records, and the context index.

Embeddings stay on Achiral at launch. Moving embeddings to a team's provider would require reindexing memory and keeping embedding spaces separate, so provider configuration affects generation but not embedding generation.

Fallback

Fallback models run on Achiral-operated infrastructure. Achiral uses them only when fallback is on and one of these is true:

• no provider is connected and the request needs a model; or
• none of your providers can serve the request, because each one is failing or keeps failing after retries.

Requests that use native tool calling through the inference gateway are not supported on your own providers yet. While a provider is connected, they fail instead of running on Achiral-hosted models.

Owners and Admins get notices when a provider incident opens or resolves.

If fallback is off and no provider can serve the request, generation fails instead of running on Achiral-hosted models. If Achiral cannot read your account's provider settings, the request fails too, rather than run on Achiral-hosted models.

Memory retrieval, deterministic app projections, and cached app narratives can still work when generation is unavailable.

Fallback billing

Fallback inference is separate from the Memory-Native Apps add-on. Accepting Memory-Native Apps terms or paying the app add-on fee does not approve automatic fallback blocks.

Each Achiral Account has a fallback allowance. After that allowance is used, paid fallback blocks can be added to the monthly bill only within the configured cap and payment rules.

Owners set the monthly block cap. When the cap is reached, or when no payment method is available, generation stops until one of these is true:

• the team connects or fixes a provider;
• an Owner raises the cap; or
• the next billing period allows more fallback use.

Fallback is measured by token cost, not request count. Achiral records the tier, provider, model, token counts, and cost for routing, audit, and billing.

Provider health

Achiral verifies a provider before marking it healthy.

Verification checks authentication, model access, and structured-output support where available.

Persistent failures mark the provider as failing and open one incident for Owners and Admins. Transient failures retry briefly; if they keep failing, the request moves to the next provider in your order and an incident opens. Providers marked failing are skipped. Achiral's fallback is used only after every working provider has been tried, and only if fallback is on.

A scheduled health job rechecks failing providers. When a provider recovers, Achiral closes the incident and notifies Owners and Admins.

Security notes

Provider API keys are encrypted at rest. Achiral never returns them in API responses.

The app shows only a key hint, such as the last four characters. Keys are not written into errors, incident reasons, run records, logs, or emails.

OpenAI-compatible endpoints must use public HTTPS addresses and pass endpoint checks. Achiral rejects URLs that contain credentials, local or cluster host names, and addresses that are private, loopback, link-local, carrier-grade NAT, multicast, or otherwise reserved. It also rechecks DNS at call time to reduce DNS rebinding risk.

Admin checklist

Before your Achiral Account starts using your providers:

1. Create a dedicated provider API key or project for Achiral.
2. Set a provider-side spending limit.
3. Connect and verify at least one provider in Achiral.
4. Decide whether fallback to Achiral-hosted models is allowed.
5. Set the monthly fallback block cap.
6. Ask Achiral to switch your account over. Achiral turns provider routing on for each account once the provider's model passes Achiral's evaluation suite.