Model providers

The catalog lists 34 providers, including vLLM, Ollama and a custom entry for any OpenAI-compatible endpoint. The API key is sealed in the vault, and moving an agent to another provider changes its model, not its instructions or tools.

A provider is connected once, at the organization level, by an admin or an owner of the organization. Its models are then available to every project. Rikoo tests the key, lists the models the endpoint actually serves, and fills in what its registry knows about each one: context window, input modalities, price.

Connecting a provider

1. In LLM APIs, click Add a provider, choose the provider in the catalog and paste its API key. For Azure, Cloudflare, vLLM, Ollama and the custom entry, also enter the endpoint.
2. Click Connect the provider. The API server seals the key in the vault at once, and the key is never shown again, to anyone.
3. The worker, the only part of Rikoo that talks to providers, tests the key and lists the models the endpoint serves. The models the catalog knows for this provider arrive switched off, and the ones the endpoint really serves are switched on.
4. Switch off the models your teams should not use. A switched-off model cannot be chosen on an agent.

Your own model

vLLM and Ollama have their own entries: give the address of your server. Any other endpoint that speaks the OpenAI API, an internal gateway for example, goes through the OpenAI-compatible (custom) entry. Its base URL must expose /chat/completions and /models, and a model the endpoint does not list can be added by hand with Add a model.

The anti-SSRF guard refuses private network addresses by default. To reach a model server on your own network, the operator allows its address range, see Infrastructure.

Main model and fallback

An agent names a main model and, optionally, Fallback models tried in order. When a call to the main model fails, on a rate limit, an outage or a timeout for example, the run continues on the next fallback instead of failing. Fallbacks must come from a different provider: a second model at the same provider does not survive that provider being down, so the Studio does not offer one and the API refuses it.

The run records which model answered and whether a fallback served, and the cost is charged to the provider that actually answered. A slower or cheaper answer is never a mystery after the fact.

Prices and the registry

Prices, context windows, embedding dimensions and parameter support come from a registry frozen into each release: 3,174 models, taken from the price list LiteLLM publishes. Rikoo reads it from its own files and never downloads it at run time. What Rikoo knows about a model therefore only changes with an upgrade, and an install without internet access reads the same figures as a connected one.

An operator can correct the token prices of a model on the LLM APIs screen. A later discovery keeps the correction.

Some models are billed per unit rather than per token: an image, a second of audio, a character of synthesized speech. The registry records the unit, and the cost of a call is computed in it. When no price is known, the cost is recorded as unknown, not as zero.

A provider unfolded: every discovered model, its context window, its price, and the switch that makes it selectable.

Where do the credentials live?

In the vault. The API server seals them with the public key of the vault's sealing key pair and cannot open them. Only the worker holds the private key: it opens a credential in the stack of an outgoing call, and each access is written to the audit log as a secret_accessed entry. See Secret vault.

Can a project use a provider the organization has not connected?

No. Providers are connected once, at the organization level, and projects choose their models from that list.