Skip to content
Guides

OpenClaw model providers

The four-page reference behind provider setup, why adding provider auth does not change your primary model, the provider and runtime split, key rotation in the control interface, and custom base URLs for local inference

6 min read

Providers here means model providers, not chat channels. The reference is an index that splits into four pages by what you are actually doing: reading the rules, configuring keys in the interface, setting up a bundled provider, or wiring something custom. Here is the map and the rules that catch people out most often.

The four pages

  • Quick rules covers model references and the command-line helpers, the rule that adding provider auth does not change your primary model, the split between a provider and the low-level agent runtime, and which runtimes are driven by a command-line tool rather than an HTTP route.
  • The control interface page covers configuring providers from the settings screen, how plugin-owned providers behave there, and API key rotation with its key sources, priority order and the conditions that actually trigger a switch.
  • Official provider plugins covers each bundled provider with its identifier, authentication variable, an example model and the quirks worth knowing, including subscription-style hosted options and the gateways that aggregate several vendors.
  • Custom providers covers everything reached through the providers section of configuration: a custom base address, a proxy, or a local inference server, along with the default optional fields and the route-shaping rules a proxy has to respect.
Reference for LLM/model providers (not chat channels like WhatsApp/Telegram).

The rule that surprises people

Adding authentication for a provider does not switch your agent to that provider. Credentials and model selection are separate surfaces: one proves who you are, the other decides what runs. That separation is deliberate, because a fallback chain, a utility model and an image model can each point at different providers, and a new login should not silently rewrite any of them. The companion surprise is the provider and runtime split: a model reference chooses a provider and a model, not the agent loop that executes the turn, and a provider prefix alone never selects an alternate runtime. Runtime overrides belong on the provider and model policy rather than on a whole agent or session, and choosing a runtime does not decide billing, because API-key and subscription credentials remain distinct.

Getting started quickly

  • Three commands cover the common path: an onboarding command with an auth choice, a command that sets the active model reference, and a command that lists what is available, with the full configuration examples living in the configuration reference.
  • Rotation lives with the keys rather than with the model: several keys for one provider are collected from the documented variables in priority order and deduplicated, and a switch happens only on rate-limit style failures rather than on any error.
  • Local runtimes are not a special case: an inference server, a proxy or an aggregator is configured as a provider with a base address, and the same optional fields and shaping rules apply as for a hosted vendor.

OpenClaw models explained explains how a reference resolves and OpenClaw model authentication how the credential behind it is stored.

Keeping the pages findable

The split kept every anchor from the previous single page, so an existing link to a specific provider section still resolves and lands on the page that now holds it. That is a small detail with a large effect: provider sections are exactly the kind of link people paste into issues and chat threads, and breaking them would scatter the answers. The same instinct shows up in the reference itself, where each provider entry keeps its identifier, its auth variable and an example model together, so a reader can copy one block rather than assembling it from three pages. OpenClaw model failover covers what happens when a provider fails and OpenClaw configuration the keys these pages write.

On Diali

On Diali the provider wiring is part of the platform rather than a per-customer setup task, with the runtime configuration generated from the dashboard and replaced at each release. Hosted OpenClaw on Diali describes the hosted assistant and Diali pricing the plans it ships in.

  • Four pages: rules, keys, bundled plugins, custom providers.
  • Adding auth never changes your primary model.
  • Local servers are configured exactly like hosted providers.
Get started

Stop reading about it, build one

Set up an agent, pick a channel, and have it working inside the app you already keep open.