n8n Anthropic nodeConfigure Anthropic in n8n.

Hand one or more documents to a Claude model and get a written answer about their content, returned as workflow data the next node can read.

• Model: required, the model that reads the document; the node will not run while it is empty.
• Text Input: the question asked about the document, for example What's in this document?.
• Input Type: url for Document URL(s), binary for Binary File(s) already carried by the workflow.
• URL(s) or Input Data Field Name(s): the addresses, comma separated, or the binary field names holding the documents.

Pushes a single file to the Anthropic API so later runs can point at it instead of resending the content every time.

• Input Type: url to fetch the file from a File URL, binary to take a Binary File from the current item.
• URL: the address of the file to upload, when the input type is a URL.
• Input Data Field Name: the name of the binary field holding the file, typically data after a download or an email attachment step.
• File Name: in Options, the name given to the uploaded file, which is what you will recognize later in a listing.

Reads back what the Anthropic API knows about one uploaded file, without touching its content.

• File ID: the identifier of the file to get metadata for, in the file_123 shape, most often taken from a previous step with {{ $json.id }}. An expression like this reads a field of the current item at run time instead of a value typed by hand.

Returns the files held by the Anthropic API, one item per file, which makes it the starting point of any housekeeping workflow.

• Return All: on, n8n keeps paging until the whole list is out; off, it stops at Limit. Paging means the API hands back the list in slices and n8n asks for the next one.
• Limit: the maximum number of files to return when Return All is off, which is the setting to keep while testing so a run stays cheap and short.

Removes one file from the Anthropic API for good, which is the operation you put behind a filter rather than at the top of a workflow.

• File ID: the identifier of the file to delete, usually piped from a listing step as {{ $json.id }} so the deletion targets exactly what the previous node selected.

Sends images to a Claude model with a question attached, and returns the description or the answer as text in the item.

• Model: required, the model doing the reading.
• Text Input: what you want to know, for example What's in this image?.
• URL(s): image addresses, several allowed when separated by commas.
• Length of Description (Max Tokens): in Options, fewer tokens give a shorter, less detailed description. A token is the unit a model counts in, roughly a word fragment.

Turns a short description of an intention into a full prompt, so you are not staring at an empty prompt field.

• Task: the description of the prompt's purpose, written plainly, for example A chef for a meal prep planning service.
• Simplify Output: on, the node returns a simplified response rather than the raw data, which keeps the next node's field mapping readable.

Takes a prompt that already exists and returns a reworked version, guided by what you say is wrong with it.

• Messages: the messages that make up the prompt to improve, added one by one with Add Message.
• Prompt and Role: the content of each message and who sends it, user to ask, assistant to set a tone or a personality.
• System Message: the existing system prompt to incorporate, if there is one.
• Feedback: what should change, such as Make it more detailed and include cooking times.

Converts a fixed prompt into a template with variables, which is what you want before running the same instruction over hundreds of rows.

• Messages: the messages that constitute the prompt to be templatized.
• Prompt: the content of the message to send, for example Translate hello to German.
• System Message: the existing system prompt to templatize, such as You are a professional English to German translator.

The main operation: it creates a completion with an Anthropic model and hands the text back to the workflow.

• Model: required, and the one field the node blocks on.
• Messages, Prompt, Role and System Message: the conversation you send, with the system message setting the behavior once.
• Web Search and Web Search Allowed Domains: enable search, then restrict it to a comma-separated domain list. It conflicts with Web Search Blocked Domains, so pick one side.
• Output Randomness (Temperature): lower it and completions get less random, close to zero they become deterministic and repetitive.