Invoke Claude, Analyze PDFs and Images, and Run Managed Agents in Jinba Flow

Use ANTHROPIC_INVOKE to get a response from Claude, ANTHROPIC_INVOKE_WITH_FILE to analyze PDFs and images, or ANTHROPIC_AGENT to run a managed agent and write results to a file with OUTPUT_FILE.

The Anthropic tool allows you to interact with the Claude AI model. It provides a simple interface to send prompts and receive responses.

• ANTHROPIC_INVOKE
  • Send prompts to Claude AI
  • Receive structured responses via JSON
• ANTHROPIC_INVOKE_WITH_FILE
  • Send PDFs and images (PNG, JPEG, GIF, WEBP) to Claude together with a prompt
  • Receive text or structured responses via JSON
• ANTHROPIC_AGENT
  • Run a Claude Managed Agent loop as a single workflow step, with file read/write/edit and web tools
  • Mounts input files for the agent and returns files the agent produces

Connect with an Anthropic API key

For further details, click here.

To get an Anthropic API key, please sign up at the official Anthropic website and follow the developer documentaiton.

Note: Treat API keys as sensitive information and never commit them to public repositories.

Get a response from Claude

- id: claude
  tool: ANTHROPIC_INVOKE
  config:
    - name: version
      value: claude-sonnet-5-5
    - name: token
      value: "{{secrets.ANTHROPIC_API_KEY}}"
  input:
    - name: prompt
      value: What is the capital of France?

Analyze a PDF or image with Claude

ANTHROPIC_INVOKE_WITH_FILE sends the files in file_link to Claude with the prompt. Claude reads charts, tables and scanned pages, not only text. Limits: up to 20 files and 20 MB in total per request; each image up to 3.75 MB and 8000 px per side. prompt is required. Temperature ranges from 0 to 1.

- id: input_file
  tool: INPUT_FILE
  input:
    - name: description
      value: "Invoice PDF or image"

- id: read_invoice
  tool: ANTHROPIC_INVOKE_WITH_FILE
  needs: [input_file]
  config:
    - name: version
      value: claude-sonnet-4-6
  input:
    - name: prompt
      value: Extract the invoice number and the total amount.
    - name: file_link
      value: "{{steps.input_file.result}}"
    - name: json_schema
      value: |
        {"type":"object","properties":{"invoice_number":{"type":"string"},"total":{"type":"number"}},"required":["invoice_number","total"]}

Extract large JSON schemas with Claude

ANTHROPIC_INVOKE and ANTHROPIC_INVOKE_WITH_FILE send json_schema to Claude as a grammar the API enforces. Anthropic rejects a schema whose compiled grammar is too large (roughly 110 fields on Sonnet 4.6; fields that also allow null lower the ceiling). The json_schema_mode config decides what happens then:

json_schema_mode	Behaviour
auto (default)	Grammar-constrained first. If the API rejects the schema as too large, the step resends the request with the schema in the prompt and validates the answer against the schema.
constrained	Grammar only. A too-large schema fails the step with the API error.
prompt	Always puts the schema in the prompt and validates the answer afterwards. Use it when you already know the schema is large.

result.schema_mode records which read produced the answer: constrained or prompt. When a prompt-mode answer is not JSON or does not match the schema, the step returns a SchemaMismatch error with the detail instead of result.content. The validator matches pattern and patternProperties with a linear-time engine (RE2 syntax): lookaround and backreferences are not available, . does not match a carriage return, and \s covers ASCII whitespace only; \uXXXX escapes work as usual. A schema using unsupported syntax, or $async, is refused as Invalid schema before the prompt-mode read is made (in auto, only after the grammar has rejected the schema).

- id: input_file
  tool: INPUT_FILE
  input:
    - name: description
      value: "Fixed-asset register PDF"

- id: read_register
  tool: ANTHROPIC_INVOKE_WITH_FILE
  needs: [input_file]
  config:
    - name: version
      value: claude-sonnet-4-6
    - name: json_schema_mode
      value: prompt
  input:
    - name: prompt
      value: Extract every field of the fixed-asset register.
    - name: file_link
      value: "{{steps.input_file.result}}"
    - name: json_schema
      value: |
        {"type":"object","properties":{"asset_no":{"type":"string"},"acquired_on":{"type":["string","null"]}},"required":["asset_no"]}

Run a managed Anthropic agent

ANTHROPIC_AGENT runs a Claude Managed Agent that can read, write, and edit files and use web tools autonomously. Input files (up to 20, passed via files) are mounted under /workspace/inputs/ for the agent, and files the agent writes to /mnt/session/outputs/ are returned as result.files — each entry has content, filename, and fileType, shaped so it can be passed directly to a downstream OUTPUT_FILE step. The agent’s final text output is returned as result.content, along with stop_reason and timed_out.

Available models are claude-opus-4-7 (default), claude-sonnet-4-6, and claude-haiku-4-5-20251001. The timeout_sec input sets a hard wall-clock timeout in seconds (60–1800, default 600). If you bring your own Anthropic API key, you must also supply your own agent_id and environment_id in config; in that case the model is fixed on your Anthropic-side Agent and the model selector is ignored.

- id: input_file
  tool: INPUT_FILE
  input:
    - name: description
      value: "CSV file for the agent to analyze"

- id: agent
  tool: ANTHROPIC_AGENT
  needs: [input_file]
  config:
    - name: model
      value: claude-opus-4-7
  input:
    - name: prompt
      value: |
        Read the attached CSV, analyze monthly sales trends,
        and write a summary report to /mnt/session/outputs/report.md
    - name: files
      value: "{{steps.input_file.result}}"
    - name: timeout_sec
      value: 900