Skip to main content

Anthropic integration

Run Claude prompts, analyze documents and images, count tokens, and inspect available models on behalf of your clients.

What it does

The Anthropic integration lets your agency call Claude models from inside any workflow. Connect a client's Anthropic API key once and your workflows can summarize incoming form submissions, extract structured fields from free-form text, classify support tickets, draft replies in the client's voice, count tokens before a request to keep spend predictable, and surface the model catalog so end users can pick the right model for the job.

Connect an Anthropic account

  1. Open your workspace in TaskJuice and navigate to Connections.
  2. Choose Anthropic and click Connect.
  3. In a new tab, open the Anthropic Console API keys page signed in as the client (or as your agency, if the client has delegated key creation to you).
  4. Click Create Key, give it a name that identifies the workspace, and copy the sk-ant-... value.
  5. Paste the key into TaskJuice and save the connection.

To rotate or revoke the key later, return to the Anthropic Console API keys page, delete the entry, create a new key, and update the TaskJuice connection.

Triggers

Anthropic does not publish a public webhook surface, so the integration is action-only per the AI-provider exception in the catalog policy. Use a Schedule, a polling trigger from another app, or an inbound webhook from a different source to invoke Anthropic actions inside a workflow.

Actions

  • anthropic/create-message sends a prompt to a chosen Claude model (for example, claude-opus-5) with a list of messages, an optional system prompt, and a max-tokens cap.
  • anthropic/count-tokens returns the input-token count for a prompt against a chosen model, so a workflow can short-circuit before sending an oversized request or branch on expected cost.
  • anthropic/list-models returns the model catalog available to the connected key, useful for surfacing a model picker to end users or for branching on model availability.
  • anthropic/get-model fetches metadata for a single Claude model by its identifier.
  • anthropic/analyze-document asks Claude a question about a PDF. Set the source input to file to send a PDF from an earlier step, or to url to send a public PDF URL.
  • anthropic/analyze-image asks Claude a question about an image, from an earlier step or from an image URL, the same way.

Reading the response

Create Message returns content as an array of blocks, not a single string. Current Claude models reason before they answer, so the first block is normally a thinking block and the reply lives in a later text block. Map the text block explicitly rather than reaching for content[0].text, which is empty on those models:

$steps.create_message.data.content[type='text'].text

usage.cache_read_input_tokens and usage.cache_creation_input_tokens are also returned, so a workflow can branch on how much of a prompt was served from Anthropic's prompt cache.

Known limitations

  • The integration calls the Anthropic REST API with the connected key. Per-key rate limits, monthly spend caps, and model-access entitlements are governed by the client's Anthropic organization, not by TaskJuice. When Anthropic returns a 429 or a 529 overloaded response, TaskJuice surfaces it as a retryable rate-limit or provider error and honors the recommended backoff.
  • Temperature is only accepted by Claude 4.6 and older models. Current models (Opus 5, Sonnet 5, Fable 5, Opus 4.8, Opus 4.7) reject the parameter with temperature is deprecated for this model, so the field is left blank by default and should stay blank unless the workflow targets an older model.
  • Streaming responses are not exposed. Each Create Message call runs as a single request and resolves only after the full response is returned by Anthropic.
  • Tool-use and prompt-caching parameters are not first-class fields on the Create Message action yet. The messages field accepts the full Anthropic content schema, so structured content blocks can still be passed through directly. To send a document or an image, use Analyze Document or Analyze Image.
  • Analyze Document and Analyze Image send a file from an earlier step inside the request. A PDF can be up to 10 MB. An image can be up to about 7.5 MB, because Anthropic measures its 10 MB image limit on the encoded image. For a larger file, use the url source with a URL Anthropic can download.
  • The integration is action-only. To react to events in real time, use a Schedule trigger, a polling trigger from another app, or an inbound webhook from a different source and call Anthropic inside the workflow.
Was this helpful?