Connecting a provider

AI Studio runs on your own provider key, connected from Settings, AI Settings, under the Provider & Model tab. Before you can chat or generate, connect at least one provider. See Bring your own AI key for the idea behind this.

Quick overview

1. Open Settings, AI Settings, and pick a provider tab.
2. Paste your API key.
3. Pick a model from that provider's list.
4. Set a reasoning effort, and optionally a system prompt, temperature, and max tokens.
5. Test the connection, then turn the provider on for chat.

The setup flow

For each provider, the steps are the same three fields, plus optional advanced settings.

Step	Detail
API key	A password-masked field. Once saved, it is shown only as a short hint, never the full value
Model	A dropdown of that provider's available models
Reasoning effort	A provider-specific set of levels, shown only for chat-capable providers
Test	Sends a small request to confirm the key works before you rely on it
Use for chat	A switch that makes this the active provider for chat

Reasoning effort options differ by provider, for example Default, Low, Medium, High, and sometimes Extra High or Max on providers that support deeper reasoning; a provider without this capability skips the control entirely. The value applies automatically to every chat request on that provider, there is no per-message toggle in the composer.

The settings screen at a glance

The Provider & Model tab shows a left-hand list of every provider. A shield icon marks each provider that already has a key saved, and the one currently powering chat is tagged Active. A summary strip across the top always tells you which provider and model chat is using right now, and whether a key is connected, so you can confirm your setup without opening a panel.

Click a provider in that list to open its panel on the right, laid out as three numbered steps:

1. API Key, a masked field. Once you save, it collapses to a Change button and a delete (trash) control, and never shows the full key again.
2. Model, a dropdown of that provider's models, with the sensible default marked recommended.
3. Try it, the Test connection button.

If the provider supports adjustable reasoning, a small inline Reasoning dropdown sits between Step 2 and Step 3. Below the three steps, an Advanced (optional) accordion holds the finer settings. Turning a provider on for chat is the Use for chat switch: you cannot flip it off directly while that provider is the active one, instead you switch a different provider on, which automatically hands it the Active tag.

Advanced options

An optional, collapsed section adds finer control.

Setting	Range	Notes
System prompt	Free text	Custom instructions layered under every reply
Temperature	0 to 1.5	Labeled Precise, Balanced, or Creative as you move the slider
Max tokens	256 to 16384	Default 2048. Caps how long a single reply can be

Each advanced field saves on its own the moment you click away from it (on blur), with a small confirmation toast, so there is no separate Save button for these three. Once connected, your key is shown masked, and you can change or delete it at any time. You can connect as many providers as you like at once, each one keeps its own key, model, and reasoning effort independently, so there's nothing to reconfigure when you switch which one is active.

Reasoning effort levels by provider

The exact set of levels differs by provider, some go further than others.

Provider	Levels
OpenAI	Default, Off, Low, Medium, High, Extra High
Anthropic Claude	Default, Low, Medium, High, Extra High, Max
DeepSeek	Default, Off, High, Max (no Low or Medium tier)
Qwen	Default (off, recommended), Low, Medium, High
Kimi	Default, Off, On (a simple on or off, no graded tiers)
GLM	Default, Off, On; a higher-end GLM model additionally unlocks Low, High, and Max on top of the basic on or off
Google Gemini	Default, Minimal, Low, Medium, High

A provider without a reasoning control at all, such as one that doesn't support adjustable effort, simply skips this field.

Step by step

The screen looks the same for every provider, so once you have done it once the rest are quick. These walkthroughs use exact clicks and locations.

Connect an OpenAI key and get ready to chat

1. Open Settings, AI Settings, and stay on the Provider & Model tab.
2. In the left-hand provider list, click OpenAI. Its panel opens on the right.
3. Under Step 1, API Key, paste your key into the masked field. If you do not have one yet, use the small external link next to the provider's name to open OpenAI's own developer console and create a key on its API keys page, then come back and paste it. Click Save.
4. Under Step 2, Model, pick a model from the dropdown, or leave the one already marked recommended.
5. If a Reasoning dropdown appears between Step 2 and Step 3, choose a level or leave it on Default.
6. Under Step 3, Try it, click Test connection and wait for the message "Connection successful, key is valid."
7. Turn on the Use for chat switch if OpenAI is not already tagged Active. That makes it the provider chat uses.
8. Optional: open the Advanced (optional) accordion to add Assistant instructions (a system prompt), nudge Creativity (temperature), or cap Max reply length. Each of those fields saves on its own when you click away, so there is nothing else to press.

Set up a video-only provider like fal.ai

1. Open Settings, AI Settings, Provider & Model tab, and click fal.ai in the left-hand list.
2. Paste your fal.ai key (formatted as a key id and secret joined by a colon) into the API Key field and click Save. You get it from the Keys page of the fal.ai dashboard, reached from the external link by the provider name.
3. Notice there is no model dropdown and no Use for chat switch here. fal.ai is a generation engine, so its card shows a Video engine badge instead, and explains that the actual model is chosen later, at generation time.
4. Clicking Test returns "Key saved (this provider has no live test)" rather than truly validating it, because this provider has no cheap endpoint to ping. That is expected, not an error.
5. Pick the model when you generate: in the composer's Video mode (Seedance 2.0, Seedance 2.0 Fast, Veo 3.1, Veo 3.1 Fast, or Kling 2.6), or inside Marketing Studio. See Generating video for the mode's controls. BytePlus works the same way, with a Video + image engine badge, and picks its model at generation time too.

Switch your active chat provider without losing your other keys

1. Open Settings, AI Settings, Provider & Model tab.
2. In the left-hand list, click a different provider that already has a saved key, shown by a shield icon.
3. Turn on its Use for chat switch. It becomes the Active chat provider immediately, and the previous one loses the Active tag on its own (you never toggle the old one off by hand).

Every other provider keeps its key, its chosen model, and its reasoning level saved, so switching back later needs no reconfiguration. Nothing you connected is lost by changing which one is active.

Connect a self-hosted or custom OpenAI-compatible endpoint

1. Open Settings, AI Settings, Provider & Model tab, and click Custom Endpoint in the list.
2. Enter a Base URL and a Model Name instead of picking from a fixed model dropdown.
3. There is no live Test connection here: the button area reads "No test for a custom provider, try it in chat" instead. Send a message in chat to confirm it works.
4. Turn on Use for chat to make it active.

A custom endpoint is the one provider whose key is not encrypted on the server and not proxied, since there is no fixed host to route through. Its key stays in your browser and requests go straight from the browser to the Base URL you typed, so use only an endpoint you trust. This is the same behavior noted under Custom endpoint below.

Your keys are safe

Keys are stored server side

Your key is encrypted at rest on the server and is never sent back to the browser. Every provider call is made from the server on your behalf, so the page itself never holds your real key, only a placeholder in its place.

Not every provider can be live-tested

Test Connection makes a real request for most providers. A couple of video-only providers do not expose a suitable endpoint for a live test, so saving the key there simply confirms it was stored, not that it works.

Which providers

Different providers support different things: some only power chat, some only generate images or video, and a couple do both. See the Provider and model matrix for the full breakdown by mode.

Custom endpoint

If your provider speaks the OpenAI-compatible chat completions format, connect it as a Custom Endpoint by giving a Base URL and a Model Name instead of picking from a fixed list. This is the way to use a self-hosted model or a less common provider.

Custom endpoints skip the server proxy

Named providers are called from the server, so your browser never talks to them directly. A custom endpoint is the one exception: since there is no built-in proxy route for an arbitrary URL, requests to it are sent straight from your browser to the address you provide, and its key stays in the browser rather than being encrypted server side.

Changing or removing a key

Return to the same provider tab any time to paste a new key over the old one, or delete it entirely. Deleting a key does not delete the models or conversations you already used with it, it just stops that provider from being usable until you add a key again. If the provider you delete was the active chat provider, switch a different connected provider on before you go back to chatting.

If a key doesn't work

A failed Test Connection almost always means one of a few things: the key was mistyped or has trailing whitespace, the key belongs to a different product from the same company, or the account behind the key has no quota or billing set up yet on the provider's own side. The error message returned by the provider is shown as-is, so check it first before assuming the app is at fault.

Troubleshooting

Reasoning effort resets on a new browser or device

Unlike your API key, which is encrypted and stored on the server so it follows you to any device you sign in on, the reasoning effort level you pick per provider is saved only in the current browser's local storage. It does not travel to a different browser, a different computer, or a private window, where it starts again at Default. If your replies suddenly feel shallower after switching machines, re-pick the reasoning level there. Your key, model, and advanced settings are unaffected, only the reasoning level is local to the browser.

What each Test connection message means

Test connection maps a raw provider error to a short, plain message, and is rate-limited to 10 tests per minute per account.

Message	What it means
Connection successful, key is valid.	The key worked and the provider answered.
Key saved (this provider has no live test)	fal.ai and BytePlus have no cheap endpoint to ping, so saving the key is all that is confirmed.
No test for a custom provider, try it in chat	A custom endpoint has no test action; send a chat message to check it.
Key is invalid, you may have copied it incompletely	The provider rejected the key (a 401 or 403 style error). Re-copy it in full.
Provider is busy (429), try again shortly	The provider returned a rate-limit (429). Wait a moment and retry.
Could not reach the provider (network/CORS)	A network or CORS failure stopped the request from arriving.
Save your key first.	There is no key saved yet to test.
Too many tests, try again in a minute.	You went over 10 tests in one minute; pause briefly and try again.

Deleting a key does not remove what you made

Removing a provider's key only stops new requests to it. Your gallery of generated images and clips, and the conversations you already had, stay in place. See Changing or removing a key above for the exact steps.

Related

• Bring your own AI key
• Provider and model matrix
• Using AI chat
• AI chat assistant for the in-editor version of this same setup