On a Hermes-edition ClawBox, the provider panel drives Hermes’ own configuration. There is no separate ClawBox copy of your provider settings: what you choose here is what Hermes uses, in chat, in Telegram and for the agent’s own work. You reach it in two places, and they are the same panel:
  • during first-boot setup, at the AI provider step — which opens on ClawBox AI, with Show more providers… for the rest and Skip — I’ll use only local AI at the bottom;
  • afterwards, in Settings → Providers, headed AI Providers.
Every connection — a device code, a sign-in, an API key, a save — runs behind a progress overlay that stays up until the box has actually applied it, so “connected” is something the box has done rather than something the screen said. In Settings the panel opens with a highlighted card naming the provider the box is running right now, the model it resolves to, and whether it is connected. Below it is the provider list, each row carrying its own connection state: choosing a row that is already connected makes it the default straight away, while choosing one that is not walks you through its sign-in or API key first.

​
Providers in the panel

Sign in

Anthropic (Claude) · OpenAI (Codex) · GitHub Copilot · Nous Portal

Paste an API key

OpenRouter · Anthropic · Google Gemini · z.ai / GLM · Kimi
Anthropic appears in both lists — sign in through Hermes’ OAuth flow, or paste an API key, whichever suits you.
The sign-in row is Anthropic, not a Claude Pro/Max subscription lane. It is Hermes’ own provider OAuth; what it grants is whatever your Anthropic account authorises. There is no claude-code row in this panel.
ClawBox AI sits above the rest as its own row. It works on this edition: it is wired into Hermes as a custom provider, so it behaves like any other provider once selected. See Subscriptions for what each plan includes.
The panel is a deliberate shortlist, not the whole of what Hermes supports. Hermes knows a much longer provider registry; anything not listed here stays reachable from the Hermes dashboard on the desktop. Configure it there and it shows up in the ClawBox panel and chat pills afterwards.

​
Connecting a provider

1

Pick the provider

Select its row. The panel loads the models that provider offers.
2

Sign in, or paste a key

For a sign-in provider, use Sign in. The flow runs inside the panel on the same origin — ClawBox drives Hermes’ own OAuth rather than sending you to the dashboard on port 8090, which is what makes it work over a remote-access tunnel. The provider’s consent page may open in a new tab; finish there and come back, and the panel picks it up.Anthropic and GitHub Copilot sign in a little differently: their login is the provider’s own attended command-line flow, so the card shows the link and the code on screen, in three plain steps, and you paste the code back into the card. The credential is minted and stored by the provider’s own tool on the device — ClawBox relays the link out and the code in, and holds nothing. If the tool is not installed, the card says so plainly rather than failing at the end, and a cancelled sign-in leaves nothing behind.For a key provider, paste the key into the API key field. The key is stored by Hermes on the device and is never displayed back to you.
OpenAI is sign-in only here. The providers that take a pasted key are exactly OpenRouter, Anthropic, Google Gemini, z.ai/GLM and Kimi — an OpenAI sk-… key is rejected by this panel. Use Sign in with OpenAI (Codex), or add the key from the Hermes dashboard, where the full registry lives.
3

Choose a default model

Pick from Default model. The list is what that provider actually offers on your account, so it is short until you have signed in or added a key.
4

Save

Save model & provider. The key, if you entered one, is stored first — it is what unlocks the model list.
Switching provider always resets the model to that provider’s default. This is deliberate: model identifiers are not portable between providers — OpenRouter spells a model one way and the vendor’s own API spells it another — so carrying the old name across would produce a device configured with a model its provider has never heard of.
If a provider has no usable model yet — not signed in, or no key — saving is refused with a message saying so rather than leaving the device pointed at something that cannot answer.

​
The on-device model

A Hermes ClawBox can also run a model locally, with nothing leaving the box. It is configured separately, in Settings → Local AI, because it works the same way on both editions.
  • It appears in the provider and chat pickers as Gemma 4 (on-device).
  • It runs with on-demand standby: the model wakes on the first request and sleeps again afterwards to give the RAM back. The first message after a quiet period is therefore slower than the ones that follow.
  • Its reasoning control is a two-state switch — Thinking off / Thinking on, not the eight-level effort scale. Its backend takes a boolean (enable_thinking) and has no graded middle, so offering Low/Medium/High would be offering settings that all do the same thing.
  • Turning it on makes it available; it does not take over. Making it your default is a separate, explicit choice.
ClawBox Connect (Jetson Orin Nano, 8 GB) runs a small model comfortably and is best paired with a cloud provider for heavier work. See Hardware.

​
Reasoning effort

Hermes takes a reasoning level per request. In the ClawBox chat the brain icon offers eight: None, Minimal, Low, Medium, High, X-High, Max, Ultra. That control is a per-conversation override — it does not rewrite the device default. Two details worth knowing:
  • Not every provider accepts every level. Choosing a provider that tops out lower clamps your selection down to the nearest level it supports rather than failing.
  • ClawBox AI shows seven — its backend rejects Ultra outright, so the level is not offered rather than being offered and failing.
  • The on-device model shows a two-state switch, Thinking off / Thinking on, instead of the eight-level scale.
  • A provider that accepts no reasoning level at all hides the control entirely.

​
Changing provider later

Any time, from Settings → Providers. Nothing needs restarting and no re-setup is involved; the next message uses the new provider.