Anthropic

View on Github

A next-generation AI assistant for your tasks, no matter the scale

Before setting up

Before you can connect you need to make sure that:

• Anthropic API token connection type: You have an Anthropic account and have access to the API keys.

• Microsoft Foundry connection type: You have a Microsoft Azure account and a model resource. Follow this guide to set up your Anthropic model. To obtaint credentials, go to Microsoft Foundry -> Playgrounds and click View code. Select Key authentication and copy your Endpoint and API key. Copy your Deployment name from the generated code.

• Amazon Bedrock (AWS Credentials) connection type: You have AWS Access key and Secret key and the Region code where the models were enabled. The IAM user must have these permissions enabled: bedrock:InvokeModel, bedrock:ListFoundationModels, bedrock:ListInferenceProfiles, and bedrock:GetInferenceProfile. When using a cross-Region inference profile, bedrock:InvokeModel must also be allowed for the profile and its destination model resources.

• Amazon Bedrock (API Key) connection type: You have a Bedrock API token generated in the AWS console under Bedrock -> Settings -> API keys, and you have the Region code.

• Google Vertex AI connection type: You have a Google Cloud project with billing and the Vertex AI API enabled, access to the required Claude models in Model Garden, and a service account with the Vertex AI User (roles/aiplatform.user) role. Download the service account key as JSON. You also need the Google Cloud location where the models are available. Use global for the global endpoint, us or eu for a multi-region endpoint, or a specific region such as us-east5.

Note: Foundation models are automatically enabled in AWS commercial regions when first invoked. First-time Anthropic users may still need to submit their use case details, and AWS Marketplace-backed models require the appropriate Marketplace permissions. Model availability and inference profiles are region-specific and can be restricted through IAM policies and Service Control Policies.

Connecting

Navigate to Apps and search for Anthropic. Select Anthropic. Click Add Connection and name your connection for future reference e.g. ‘My connection’.

Anthropic API token

1. Fill in your API key. You can create a new API key under API keys. The API key has the shape sk-xxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxx.
2. Click Connect.

Microsoft Foundry

1. Fill in your API Key, Endpoint and Deployment name.
2. Click Connect.

Amazon Bedrock (AWS Credentials)

1. Fill in your Access key, Secret key and Region.
2. Click Connect.

Amazon Bedrock (API Key)

1. Fill in your API token and Region.
2. Click Connect.

Google Vertex AI

1. Fill in your Google Cloud Project ID.
2. Set the Location where the Claude models are available.
3. Paste the complete service account key JSON into Service account JSON.
4. Click Connect.

The model dropdown contains the supported Google Vertex AI model IDs. Model availability depends on the selected location and the models enabled for your Google Cloud project.

Actions

Note: Only the ‘Anthropic API token’ connection type supports skills.

Chat actions

• Chat action has the following input values in order to configure the generated response:

1. Model (Available direct models and active system-defined or application inference profiles are listed in the dropdown)
2. Prompt
3. Max tokens to sample
4. Temperature
5. top_p
6. top_k
7. System prompt
8. Stop sequences

For more in-depth information about action consult the Anthropic API reference.

Translation

• Translate translate file content retrieved from a CMS or file storage. The output can be used in compatible Blackbird interoperable actions. Use Filter glossary to include only glossary entries found in the source text; disable it to include the entire glossary.
• Translate text given a text and a locale, tries to create a localized version of the text.

Editing

• Edit Edit a translation. This action assumes you have previously translated content in Blackbird through any translation action. Only looks at translated segments and will change the segment state to reviewed. Use Filter glossary to include only glossary entries found in the source text; disable it to include the entire glossary.
• Edit Text given a source segment and translated target segment, responds with an edited version of the target segment taking into account typical mistakes.

Review

• Review reviews a translation. This action assumes you have previously translated content in Blackbird through any translation action. Adds a quality score to each unit in the file. Segments scoring at or above Score threshold (0 to 1, default 0.8) are set to final; segments below it are counted as under threshold. Use Filter glossary to include only glossary entries found in the source text; disable it to include the entire glossary.
• Review text reviews a single text and returns a quality score.

Note: Files are interoperable, under the hood Blackbird may convert to and from bilingual files with metadata but all our Actions are able to accept a lot of different file formats.

Background actions

Background actions run the same translation, editing and review as the actions above through Anthropic Message Batches, at about half the price. Most batches finish within an hour; a batch that does not finish within 24 hours expires.

Note: Message Batches are only available for the API Token connection type, not for Amazon Bedrock, Microsoft Foundry or Google Vertex AI.

• Translate in background, Edit in background and Review in background take the same inputs as Translate, Edit and Review (except Skill ID), submit the batch and return its Batch ID and a Transformation file. When the file has nothing to process, no batch is created and the Batch ID is empty.
• Download background file applies the results of a finished batch to the transformation file. The Batch ID is optional, because the file carries it. Requests that failed, expired or were canceled leave their segments unchanged; they are counted as Unresolved segments and the reasons are logged. For review jobs the output includes the Average score (0 to 1).
• On background job finished (event) fires once when the batch ends: completed, failed, expired or canceled. With an empty Batch ID it fires on the first check.

Token usage

For all actions you can configure the ‘Max tokens’ optional input. This value limits the number of tokens generated in the response. If left empty, the default value will be the maximum number of tokens allowed by the model. For example, Claude Sonnet 4 has a maximum of 64,000 output tokens - leaving this field empty means 64,000 will be used as the value. To limit the tokens generated in the response, set this value to a lower number.

Temperatures

Anthropic allows you to set the temperature value for and Action you perform in Blackbird. You can customize this value (between 0.0 and 1.0) but we’ve also added handy labels that are available in the Blackbird UI:

• 0.1 Governed: Maximum predictability.
• 0.3 Balanced: A controlled blend of accuracy and flexibility.
• 0.5 Expressive: Accuracy is preserved, structure is looser (default).
• 0.6 Exploratory: Encourages alternative phrasings and ideas while remaining context-aware.
• 0.8 Experimental: Prioritizes creativity over predictability.

Model Comparison

The following table compares the key characteristics of Claude models:

Model	Max Output	Context Window
Claude Opus 4.1	32,000 tokens	200K tokens
Claude Opus 4	32,000 tokens	200K tokens
Claude Sonnet 4	64,000 tokens	200K / 1M (beta) tokens
Claude Sonnet 3.7	64,000 tokens	200K tokens
Claude Haiku 3.5	8,192 tokens	200K tokens
Claude Haiku 3	4,096 tokens	200K tokens

Max output is the most important metric to consider when choosing a model for translation tasks, as it determines how much text can be generated in a single response. The context window is also important, especially for tasks that require understanding of larger documents or conversations like glossaries.

Feedback

Do you want to use this app or do you have feedback on our implementation? Reach out to us using the established channels or create an issue.