Models & providers

Where models are configured

Models live in Settings › AI Providers. What you see there depends on your role and edition:

The Models tab is the single source of truth for every chat and agent model the platform exposes. Generation models (image, video, sound, speech, music) are handled separately, see Studio.

Platform API-key providers

On the API Keys tab an admin saves one key per provider. Its models then become available wherever a model is picked:

A key saved here takes priority over the server's environment configuration; with no saved key, the server falls back to its environment. Saving a key refreshes the model list right away.

Run on your own provider key

In Your keys, paste a key for a provider and press Save key. The provider checks the key when you save it, and a rejected key shows the provider's reason. Then turn on Use my key for that provider. Each row states its route: Runs on your key, Runs on the LiveContext key, or Runs on the LiveContext key (yours is saved). You can switch back at any time, and your key stays saved. OpenRouter and Cohere are not offered here.

On your key, the provider bills you the tokens directly. LiveContext charges a flat fee per agent turn, by the model's price band, and never more than the same turn would have cost on the LiveContext key. These turns appear as Your key in your credit history. Press See the price per turn in the panel to see the current fees.

The fee for each price band is listed under Own-key fee in Billing.

CLI bridge providers

The five bridge providers run a coding-agent CLI on a bridge host instead of calling an API directly: claude-code, codex, gemini-cli, mistral-vibe, and cursor. Claude Code and Codex let you use a Claude or ChatGPT subscription instead of an API key, Gemini CLI logs in with a Google account, Mistral Vibe uses a Mistral API key set on the bridge host, and Cursor uses a Cursor account (a Free Cursor plan only runs the Auto model; the other Cursor models need a paid Cursor plan). Each bridge, picked in the toggle of the CLI tab, has a setup panel (install the CLI, log in, start the bridge, then Verify Connection) and an access panel titled Who can use this bridge.

Max requests per user per day caps each user (leave it empty for unlimited), and Usage today lists the day's activity. Because the default is Admin only, an admin testing a bridge model sees it work while every other user is refused until the mode changes.

An API model is usable the moment its key is saved. A bridge model also needs the CLI to be installed and logged in on the bridge host. Before a bridge provider is offered in a model picker, the bridge is asked for its status (the answer is cached for 60 seconds). A CLI that is missing or logged out is hidden, and if the bridge cannot be reached at all, every bridge provider is hidden rather than guessed at. Regular API providers are never affected.

Reasoning effort

Reasoning effort is one dial with six levels, from lightest to most thorough:

minimal, low, medium, high, xhigh, max.

Only three providers honor it: claude-code, codex, and the direct anthropic API. Gemini CLI and Mistral Vibe expose no usable effort control, and Cursor puts the effort in the model name itself, so the selector is not shown for them, nor for any other provider. Each of the three maps the level to its own setting:

Precedence when a run executes: a per-conversation or per-run override beats the per-agent setting, which beats the per-model admin default (the Effort column of the Models tab). If none is set, nothing is sent and the provider's own default applies. In the Effort column, - means “no default, let the model decide”.

Vision and attachments

Images and files you attach, and images returned by tools (for example a file an agent opens), are sent inline to the model, and a vision-capable model can see the images. Inline content has a size cap per item:

Past the cap, the model receives the file's extracted text when there is some, otherwise a placeholder naming the file: it knows the file exists but cannot read its contents. Tool-result images reach the model on both API providers and CLI bridges.

The admin model catalog

The Models tab (Model Configuration) is a sortable, per-row editable list. Its columns include Provider, Tier, Effort, the Recommended star, Price ($/1M) (USD per 1M input and output tokens), and global rate limits (TPM global, RPM global).

Deleting a regular model disables it. Deleting a custom model removes it, since there is no catalog row to fall back to. Per-tenant rate-limit columns exist but stay hidden.

When a model is disabled

Disabling a model does not break the agents, workflow nodes, and chats that already use it. On the Chat / Agent tab a disabled model shows a Replaced by selector: while the model stays disabled, those runs use the chosen replacement instead (and are billed for it). Platform default means the model the platform uses when none is chosen.

Free-plan models (cloud)

On the cloud each row has a Free chip. It decides whether a Free-plan account may spend its monthly credits on chat and agent turns with that model. It is off by default: a Free account then needs a top-up to run the model. If no model on a tab is open to the Free plan, the tab shows a warning, because Free accounts are then refused for that kind of turn.

What ships to self-hosted installs (cloud)

The cloud catalog also carries a CE: auto / CE: on / CE: off chip per row. It decides what the model catalog bundle ships to self-hosted installs, independently of the enabled toggle: auto follows it, on ships the model enabled, off ships it disabled.

Model categories

The Models tab has one pill tab per category, and a model can be enabled in one and disabled in the other:

Bridge (CLI) providers are hidden from the Browser Agent tab, where ranking or disabling them would have no effect. They still appear on Chat / Agent.

Model execution links

An execution link separates what you are billed for from what actually runs. It maps a billed provider and model to a different execution provider and model, a CLI bridge or another API provider. Credits are still charged at the billed model's price. Leaving the execution model blank reuses the billed model id, shown as same as billed.

Resolution checks the exact scope first, then falls back to ALL. Sub-agent runs, the browser agent, and other callers without a scope are reached only by an ALL link. There is at most one link per billed model and scope, and each link can be enabled, disabled, or deleted.

On the Models tab, a model whose exact id is also served by a CLI (Anthropic to Claude Code, OpenAI to Codex, Google to Gemini CLI) has a route button that creates an ALL link to that CLI in one click. Once routed, the button becomes a badge that opens a surface picker. The button turns amber when that CLI cannot run on the bridge host (missing or logged out): runs routed to it would fail until the CLI is fixed or the link is changed. Remember the bridge access policy too: with the default Admin only mode, other users' runs of a routed model are refused.

Self-hosted (CE)

LLM source: Cloud or API keys

On a self-hosted install, Settings > Cloud has an LLM source switch once the install is linked. The API Keys tab shows which source is active:

CLI bridges have no such switch: they always run on your own bridge host. See Self-hosting to link a cloud account.

The model catalog on a self-hosted install

Each release of the Community Edition ships with the cloud's signed model catalog of that day, applied when the install starts. So a never-linked install still gets the models of its release. A linked install also syncs the latest signed catalog on startup and about every 15 minutes, and the Model catalog bundle control (Update bundle) syncs on demand. The signature is verified before anything is merged, and your own edits and custom models are kept.

Troubleshooting

Related pages