Add your AI key

After this guide, your agents run on your own AI account: Anthropic, OpenAI, AWS Bedrock, Azure OpenAI, or an endpoint that works like OpenAI.

Who can do it:

• A key for every product: a workspace admin or owner.
• A key for one product only: a workspace admin or owner, or a person whose role in that product lets them change the product's own keys and tools. See Roles.
• Every plan needs a key, also Free. Your AI provider bills you for what the agents use. The plan does not change what the provider allows you to use. See Plans and limits.

Before you start

A key is the secret text that your AI provider gives you, so that a program can use your account. The sections below say where each provider makes one.

The platform encrypts a key as soon as it arrives. Only the platform can open it. The agents never see it, and nobody can read it back on the page.

Add the key

1. Open Settings → Integrations.
2. Find the group Your AI account. Choose the row of your provider and press Connect.
3. Paste the key into the field. For AWS Bedrock, Azure OpenAI and an OpenAI-compatible endpoint, the key is a short JSON text: copy the example from the section of your provider below and put in your own values. You can also press Or choose the key file and pick a file.
4. Under Use it for, choose where the key works. Every product (each product without its own key uses it) is one key for the whole workspace. "Only" and the name of a product is a key for that product only.
5. Press Save and test.

If you linked Telegram, the bot may ask you to confirm the change. The page then says Waiting for your OK in Telegram. Check that the bot's message shows the same fingerprint as the page, then tap ✅ in the bot. A fingerprint is a short code that names the key without showing it.

What you see when it worked

• The row of your provider says Connected.
• After the test, the row says Working. If the test failed, it says Did not work with the reason, and the top of the page lists it under "needs attention".
• The provider your products run on is marked In use.

A key for every product, or for one product

A key for every product is shared. Each product without its own key uses it.

A product's own key comes first for that product. To give a product its own key:

1. Open Settings → Integrations and press Manage on the provider's row.
2. Find the product in the list and press Set its own.
3. Paste the key and save it.

To change a key later, press Manage, then Replace on its line. To stop using it, press Remove. A product that loses its own key uses the shared key again, if there is one.

Which provider a product uses

A product can run on one provider only. By default, it uses the first provider that has a key, in this order: Anthropic, AWS Bedrock, OpenAI, Azure OpenAI, OpenAI-compatible endpoint.

To choose one yourself:

1. Open Settings → Products, and press Settings on the product's row.
2. Open Technical settings at the end of the page. It is closed at first.
3. Under Analytics and AI, choose the provider in AI provider. Automatic: the first provider with a key is the default.
4. Press Save in the bar at the bottom of the window.

Models for each kind of job

A job is one piece of work by an agent, for example one review or one change. You can choose the model for each kind of job.

1. In the product's Technical settings, under Analytics and AI, open Models per kind of job.
2. Type the model name for a kind of job. An empty field uses the provider's default.
3. Press Save.

Use a model that your provider account may use. The sections below say what to type for AWS Bedrock and Azure OpenAI.

Anthropic

1. In the Claude Console (once called the Anthropic Console), open Settings → API keys and make a key. It starts with sk-ant-api. The provider's guide to API keys shows how.
2. Paste it into the API key field of the Anthropic row.

Use an API account that may use the model you choose. A key from a Claude Pro or Max plan is a different thing, and it does not work here.

OpenAI

1. On platform.openai.com, open API keys and make a key. A project key limited to the Responses API is enough. A service account key of a project works too, for a key that is not one person's. See the OpenAI quickstart and the guide to API keys.
2. Paste it into the API key field of the OpenAI row.

Choose an OpenAI model that your OpenAI project may use. The limits of your OpenAI account still apply.

AWS Bedrock

The key for AWS Bedrock is this JSON text. Put in your own values:

{
  "region": "us-east-1",
  "access_key_id": "YOUR_ACCESS_KEY_ID",
  "secret_access_key": "YOUR_SECRET_ACCESS_KEY"
}

1. In AWS, choose a region and a Claude model or inference profile that your account may use. See Bedrock model IDs and regions.
2. Make an access key for a user that may only call bedrock:InvokeModel and bedrock:InvokeModelWithResponseStream.
3. Paste the JSON, with your values, into the AWS access field of the AWS Bedrock row.
4. In the product's Models per kind of job, type a Bedrock model ID for each kind of job. A Bedrock model ID for Claude contains anthropic., such as eu.anthropic.claude-….

You can add "session_token" for a key that works for a short time only. Replace that key before it expires.

Azure OpenAI

The key for Azure OpenAI is this JSON text. Put in your own values:

{
  "endpoint": "https://your-resource.openai.azure.com",
  "api_key": "YOUR_AZURE_API_KEY"
}

The endpoint is the address of your resource only, as in the example. The platform adds the rest of the address.

1. In your Azure resource, deploy a model that supports the Responses API. See Azure endpoints and deployments.
2. Paste the JSON, with your values, into the Connection field of the Azure OpenAI row.
3. In the product's Models per kind of job, type your deployment name, not the model's name.

This connection uses an API key. It does not use Microsoft Entra sign-in.

OpenAI-compatible endpoint

This is for a service or a model you host that works like OpenAI's Responses API, such as a gateway.

The key for an OpenAI-compatible endpoint is this JSON text. Put in your own values:

{
  "base_url": "https://your-provider.example/v1",
  "api_key": "YOUR_PROVIDER_KEY"
}

The base_url is an https address that ends in /v1. You can add "model" to choose the endpoint's default model.

1. Paste the JSON, with your values, into the Connection field of the OpenAI-compatible endpoint row.
2. Test the endpoint before you let agents start by themselves.

The endpoint must support the Responses API, with the tools that agents use to write code. An endpoint that has only Chat Completions is not enough.

More

• Integrations: every integration and where its key can be.
• Product settings: the other settings of a product.
• When something does not work: for example, no key, or a key the provider refuses.