Which providers and models can I connect through Manifest?

Updated

Manifest connects AI agents to four categories of model sources: bring-your-own-key (BYOK) providers, subscription plans via OAuth tokens, local model servers, and custom or OpenAI-compatible endpoints. All of them sit behind a single OpenAI-compatible endpoint, so your agent code points at one URL and Manifest handles the routing. The right mix depends on whether you want cloud models, subscription capacity, local inference, or some combination of all three.

The four connection types

Bring your own key (BYOK)

Manifest accepts API keys from popular providers. You keep your existing provider accounts and pay those providers directly for usage; Manifest routes requests and tracks cost per key, agent, and provider. This is the simplest path if you already have provider accounts and want one endpoint in front of them.

Subscription plans via OAuth tokens

If you pay for subscription-based access rather than per-token billing, Manifest supports OAuth tokens for those subscriptions. This lets you route agent traffic through plans you already pay for instead of opening new pay-per-token accounts. The site lists "subscriptions, pay-per-token, local models, and custom providers" as the supported connection modes.

Local models

Manifest routes to Ollama, LM Studio, llama.cpp, or any local server. Because routing is per-route, you can decide which requests go to the cloud and which stay on your machine — for example, sending sensitive prompts to a local model while routine calls go to a hosted provider.

Custom providers and OpenAI compatibility

Manifest is OpenAI-compatible and supports custom providers, so any endpoint that speaks the OpenAI API shape can be added. That covers self-hosted inference servers and internal gateways that aren't on the standard provider list.

Choosing what runs where

Need Connection type What you control
Use existing provider accounts BYOK Keys per provider; pay usage directly
Use subscription capacity OAuth tokens Which plans serve which routes
Keep data on your infrastructure Local models (Ollama, LM Studio, llama.cpp) Per-route cloud vs. local split
Connect an internal or non-standard endpoint Custom provider / OpenAI-compatible Endpoint URL and model parameters

Routing decisions are made per route, and model parameters can be set visually per route without changing code. That means the cloud-versus-local split is a configuration choice, not an architecture rewrite.

What happens when a provider fails

Provider choice interacts with reliability. Manifest's fallback feature automatically retries on another provider when one goes down, and Autofix repairs broken requests before the error reaches your agent. The site cites Datadog's State of AI Engineering for the claim that 5% of LLM API calls fail in production and that most of those errors are avoidable. If you connect multiple providers, fallback has somewhere to go; a single-provider setup has no alternate route.

Practical setup notes

  • Self-hosting: the full LLM router can run on your own infrastructure or machine, described as free and open source. This matters if you want local routing without sending traffic through a hosted gateway.
  • Observability: full request and response body logs plus cost visualization broken down by agent, key, and provider help you verify which connection type is actually serving traffic.
  • Pricing: Manifest has a pricing page and an enterprise contact option, but the available material doesn't specify plan costs or whether any tier requires login. Check the pricing page before assuming a free tier covers your volume.

Bottom line

Start with BYOK if you already have provider keys, add OAuth subscription tokens if you want to use existing plans, and add a local server if any traffic must stay on-premises. Registering more than one provider also enables fallback, which is the main reliability benefit of routing through a gateway rather than calling providers directly.