AI Vision
AIAnalyze images through a single connector that routes to ten vision providers, including OpenAI, Anthropic, Google, and Mistral. Describe scenes, read text, and answer questions about any image inside a flow.
What This Integration Enables
AI Vision gives an agent sight as a callable step. Point it at an image from FlowRunner storage or a URL, ask a natural-language question, and get the answer as text. Because one connector fronts ten providers, the model is a configuration choice rather than an integration project, which matters when a provider changes pricing or a newer model reads your images better. The value is not the model call itself. It is that the image analysis becomes one node in a larger flow: the answer feeds a validation step, a record update, and a decision about whether a person needs to look. An orchestration layer is what turns a vision API into a repeatable operational step, and FlowRunner is built for that layer.
Without FlowRunner
With FlowRunner
Use Case Scenarios
Damage Assessment From Field Photos
A field technician uploads a photo of damaged equipment. The agent sends it to Analyze Image with a prompt asking for the visible damage, any asset labels, and a severity read. The model returns a description and the readable asset tag. The agent matches the tag to the equipment record and attaches the analysis to a maintenance ticket. What used to require a person to open the image and transcribe it now lands as a structured summary the ops team reviews.
Document Scan Triage
Scanned forms arrive as images that a text parser cannot read cleanly. The agent passes each scan to a vision model with a prompt to extract the header fields and flag anything illegible. Clean scans flow straight into the record system. Scans the model marks as unclear are held for a person, so bad data never propagates downstream.
Content Moderation Pre-Check
User-submitted images need a first-pass review before they appear publicly. The agent runs each image through a vision model with a prompt describing what to flag. Clearly safe images publish automatically. Anything the model marks as borderline goes to a moderator with the image and the model's read attached, so a person makes the call on the edge cases.
Human-in-Loop Highlight
A vision model gives an opinion, not a verdict. When AI Vision reads an image as damaged, unsafe, or unclear, that read is an input to a person, not a decision the flow makes on its own. The agent formats the model's answer and routes it through a [human-in-the-loop](/concepts/human-in-the-loop/) step: the flow pauses, assembles the image and the model's description, sends it to the right reviewer via Slack or email, and resumes once they respond. The reviewer sees exactly what the model saw. They decide what happens next.
Agent Capabilities
2 actionsImage Analysis
2- Analyze Image Analyzes one or more images using a selected AI vision provider and model. Supports 10 providers including OpenAI, Anthropic, Google Gemini, Mistral, and others.
- Analyze Image with Structured Output Analyzes images and returns a structured JSON response matching a provided JSON Schema. Ideal for extracting specific data points, classifications, or structured information from images.
Start building with AI Vision
$100 in credits. No card required. Connect in minutes.