Vision Node

The Vision node turns images into something a workflow can act on — text, structured data, captions, classifications, or a transformed image. Pick an operation (what you want done) and a provider (where it runs), then wire an image into the node.

Quick start

1. Drop a Vision node onto the canvas.
2. Wire an image upstream — from a File node, an Image node, a screenshot, a Webcam node, or a previous Vision step.
3. Open the config panel and choose an operation. The dropdown is grouped:
   • Recognize — pull literal content out of pixels (text, tables, charts)
   • Reason — open-ended interpretation (describe, Q&A, extract, classify, moderate)
   • Modify — image-to-image transforms (background removal, enhance, filters)
4. Pick a provider if the operation supports more than one. On-device Apple is the default wherever it works.

Providers at a glance

Apple gives you privacy and cost-zero on-device processing. Cloud models give you reasoning and handwriting (billed in credits). CLI tools sit in between — local but powered by frontier models.

Operations

Recognize

Reason

Modify

Example workflows

1. Receipt scanner

Drop a photo of a receipt, get a row in your expense spreadsheet.

File ──▶ Vision (Extract, schema: { vendor, date, total, currency })  ──▶ Sheet

Set the schema field to { "vendor": "string", "date": "string", "total": "number", "currency": "string" }. The Vision node returns output.data ready for the Sheet to append.

2. Translate a foreign menu

Tourist mode in three nodes.

File or Webcam ──▶ Vision (Translate Text in Image, target: English) ──▶ Note

The Note shows the translated text. Swap the target language to French, Japanese, anything.

3. Screenshot to React component

Paste a UI mockup, get a JSX component back.

Image (paste screenshot) ──▶ Vision (Mock → React Component, provider: CLI) ──▶ Code (write file)

The Code node drops the generated component into your project. Run it again with a tweaked mockup and iterate.

4. Auto-tag a photo library

Folder (iterate images) ──▶ Vision (Tag / Keyword) ──▶ Sheet (path + tags)

The Tag operation returns labels: [{ label, confidence }]. The Sheet node can flatten that into a comma-separated cell.

5. Free, local receipt OCR at scale

Layer Apple's on-device OCR with cloud reasoning. Apple does the heavy lifting; the cloud only sees plain text.

Folder ──▶ Vision (OCR, provider: Apple) ──▶ Agent (clean + dedupe) ──▶ Sheet

6. Content moderation gate

Upload ──▶ Vision (Content Moderation) ──▶ Condition (block if flagged) ──▶ Publish

Vision returns { flagged, categories: [{ category, score }] }. The Condition node routes based on flagged.

Inputs and outputs

Input

The Vision node looks for an image on its incoming edge, checking — in this order — value, image, __image__, __chart__, chart, imageData, processedImage, output. Any upstream node that publishes one of those fields with a base64 image (or data URL) will feed the Vision node correctly.

Output

Output shape depends on the operation:

• OCR / Translate / Caption / Q&A → { text, lines?, confidence? }
• Extract / Chart / Table → { data: <parsed JSON> }
• Classify / Tag → { labels: [{ label, confidence }] }
• Count → { count, target }
• Moderate → { flagged, categories: [{ category, score }] }
• Mock → HTML/JSX → { code, language }
• Modify ops → { image, imageData, processedImage }

Output shape is the same regardless of provider for a given operation — switch from Apple OCR to GPT-4o OCR and downstream nodes don't need to change.

Forwarding images downstream

When you wire a Vision node into an Agent node, the Agent automatically picks up the image (if any) and forwards it to its multimodal model as a real attachment — not as base64 text in the prompt. This means a Vision node doing background removal can hand its result straight to an Agent for further analysis with no glue.

Tips

• Privacy-sensitive workloads → Apple. Faces, IDs, medical images: keep them on-device.
• Handwriting → cloud or CLI. Apple's OCR is print-leaning; modern VLMs handle cursive much better.
• High volume → CLI. If you're tagging thousands of photos, the CLI provider amortizes through your existing tool subscription.
• Structured extraction → write a schema. A two-line schema hint reliably shapes the JSON.
• Multiple steps are cheap. It's common to chain two Vision steps — for example "auto-crop document" → "OCR the crop."

Related

• Image Node — feed images into the Vision node
• Agent Node — receives Vision output (text or images) for further reasoning
• Sheet Node — common destination for OCR / extract output
• Apple Intelligence — what's on-device, how to check availability