Add your own AI key

Varly runs on your own AI key. That means the AI work is billed by the provider to your own account, at their prices, and Varly adds nothing on top. The key belongs to the organization, so in an agency account it works for all clients. The AI only ever sees the research of the site it is working on, never data from your other clients.

What the AI key unlocks

Without a key, Varly still finds ideas with its fixed rules. With a key, you also get:

• Smarter ideas. The AI combines all research of a site (heatmaps, click impact, analyses and survey answers) into ideas, each with the research behind it. See Ideas.
• A/B test variants. Building a test from an idea, or from changes you describe in words, needs an AI key. Without one you can still change your page yourself with Edit directly in the visual editor. See Create an A/B test.
• AI summaries of heatmaps, session recordings and survey results.
• A daily research round, if you leave it switched on (see below).

Get a key from your provider

Provider	Where to create a key	What the key looks like
Anthropic (Claude models)	console.anthropic.com, under API keys	Starts with sk-ant-
OpenAI (ChatGPT models)	platform.openai.com, under API keys	Starts with sk-
Google Gemini	aistudio.google.com, under Get API key	Often starts with AIza
DeepSeek	platform.deepseek.com, under API keys	Starts with sk-
Mistral	console.mistral.ai, under API keys	A long code of letters and numbers

Most providers ask you to add credit or a payment method before the key works.

Add your key

1. Create an API key with your provider (see the table above).
2. In Varly, open Organization settings. In an agency account, click Settings at the top of your list of clients. For a single shop, click Settings at the top of the page with your sites. You can also go to Account → Organizations and click Settings next to the organization.
3. In the AI key card, pick your AI provider.
4. Paste the key into the API key field.
5. Choose a Model.
6. Click Test the key if you want to check it first. Varly makes a tiny request to your provider (it costs a fraction of a cent) and shows "The key works." or what went wrong.
7. Tick or untick Daily research round.
8. Click Save.

Varly also checks the key with your provider before saving it. If it works, you see "Saved. The key works." The key is stored encrypted and never shown again. Afterward, the field shows the last characters of the saved key.

Choose a model

Each provider has a recommended model, which Varly picks for you. You can switch at any time; the next AI task uses the new one.

Provider	Models in the list
Anthropic	Claude Opus 5.5 (recommended), Claude Sonnet 5.5 (faster, about half the price), Claude Haiku 5.5 (cheapest)
OpenAI	GPT-6.1 Sol (recommended), GPT-6 Astra (most capable, most expensive), GPT-6 Luna (cheapest)
Google Gemini	Gemini 4 Argon (recommended), Gemini 3.8 Flash (faster and cheaper)
DeepSeek	DeepSeek V4 Pro (recommended), DeepSeek Flash (faster and cheaper)
Mistral	Mistral Large 4 (recommended), Mistral Small 4 (faster and cheaper)

The card shows the provider's price per million tokens for each model. For DeepSeek, the price depends on the time of day, so Varly shows no price there.

Another model. To use a model that isn't in the list, choose Another model and type its name exactly as your provider writes it, for example gpt-6-luna. Click Test the key to check that your key can use it. Varly doesn't know the price of such a model, so it counts tokens but shows no cost for it.

What works with this model. Below the models, the card lists the AI features and whether they work with the chosen model. All of Varly's AI features work with every model in the list. The card also says whether the model can read images. If a feature ever needs a model that reads images and yours can't, Varly says so instead of failing.

Daily research round

When Daily research round is on, the AI looks at the new research of every site in the organization once a day and adds ideas. This only happens for sites with tracking on. What a round costs depends on the model and on how much research a site has. The AI usage, last 30 days card shows an estimate.

When it's off, the AI only works when someone asks for it, for example by clicking Find new ideas on the Ideas page, building a test, or asking for a summary.

How to read the costs

The AI usage, last 30 days card shows, for the whole organization:

• Research rounds: daily and manual rounds.
• Tests built: variants the AI built.
• AI summaries: summaries of heatmaps, recordings and surveys.
• Ideas by AI: ideas the AI added.
• Estimated cost: estimated from token counts and list prices. Your provider's invoice is the final amount.

If you used a model whose price Varly doesn't know (DeepSeek, or a model you typed yourself), the card says how many tokens are not included in the estimate. Your provider's dashboard shows what they cost.

Replace or remove the key

• Replace: paste a new key into the API key field and click Save. Leave the field empty to keep the current key while you change the model or the daily round.
• Switch provider: pick the other provider, paste a key from that provider and click Save. The new key replaces the old one.
• Remove: click Remove AI key. AI ideas, test building and summaries stop until you add a key again. Rule-based ideas keep working.

Common problems

"That does not look like an Anthropic API key (sk-ant-…)". You pasted something else, or only part of the key. Copy the full key again from console.anthropic.com.

"This looks like an Anthropic key." You picked a different provider than the one the key is from. Pick Anthropic, or paste a key from the provider you picked.

"The key is not valid." The key was deleted, mistyped or only partly copied. Create a new key with your provider and paste it.

"Your … account has no credit left." Add credit, a payment method or a higher spending limit in your provider's console, then click Test the key again.

"… does not know the model". The model name is mistyped, or your key can't use that model. Check the name with your provider, or pick a model from the list.

"… is limiting requests right now." Your provider is limiting how many requests your key can make. Wait a minute and try again.

"This feature needs a model that can read images." The chosen model can't read images. Pick a model that can; the card shows which ones do.

"Only an owner can change the AI settings." Your role is Editor or Viewer. Ask an owner to add the key. See Team roles.