Segmind
AIRun serverless GPU inference across more than 200 generative AI models covering image, video, audio, and text, billed per use. Agents generate media without provisioning or paying for idle hardware.
What This Integration Enables
Most generative model connectors hand you a job ID and a polling loop. Segmind hands you the output. Its models are synchronous, so Run Model returns the result on the same request, and a flow that generates media stops needing a wait step, a status branch, and a timeout policy. For anything under a few seconds of compute, that difference is the entire ergonomics of the flow.
The second thing Segmind does that its neighbors do not is publish the economics per model. Get Model Schema returns a model's endpoint, category, whether it responds synchronously, its average cost, its average latency, and the complete list of input fields with types, defaults, and allowed values. List Models surfaces cost and latency across the catalog so a flow can shortlist before it commits. That is the difference between choosing a model on a vendor's adjectives and choosing it on numbers the vendor publishes, which is a decision an operations owner can actually defend in a budget review.
The catalog itself runs past 200 models across image, video, audio, and text, all billed per use with no idle hardware to pay for. Binary output from image, video, and audio models is validated against known file signatures and written into FlowRunner file storage, so what comes back into the flow is a URL rather than a blob you have to deal with. Text models return their JSON directly. Because each model takes a different input set, the intended pattern is schema first, then run.
Without FlowRunner
With FlowRunner
Use Case Scenarios
Listing imagery that keeps a consistent look
Raw supplier photos arrive inconsistent. The agent calls Get Model Schema once to learn the exact input fields for the chosen background removal and relighting models, then runs each image through Run Model and receives finished files back inline. The approved set is written to the product record in Shopify. Because the calls are synchronous, the whole treatment is one linear branch, and a failure is visible at the step that caused it.
Choosing the model on published economics
A team needs short video clips for a campaign and has two candidate models. Rather than argue, the agent pulls both schemas, reads the published average cost and average latency for each, multiplies by the item count, and posts both projections side by side. The cheaper model wins for the long tail of the catalog and the better one is reserved for the hero items. The split is a rule in the flow, revisitable next quarter when the published numbers change.
A model catalog that stays current without a redeploy
The model picker reads the directory Segmind publishes, so newly added models appear without any change to the connector. A flow can therefore select a model by slug at runtime, filtered by modality or search term through List Models, and adopt a better model as soon as it ships. The flow does not need to know the model's name in advance, only the shape of the job it wants done.
Human-in-Loop Highlight
The gate on Segmind is the batch cost decision, and it is only possible because the numbers exist. Before a large run, the agent calls Get Model Schema for each candidate model, reads the published average cost and average latency, multiplies by the item count, and stops there. It does not pick the cheaper model to be safe or the expensive one to be thorough, because that tradeoff is a business decision wearing a technical costume. It posts both projections to the owner in Microsoft Teams: "412 images to treat. Model A projects at the lower published cost with a longer average latency; Model B projects at roughly three times the spend and returns faster. Which one, or split by product tier?" The person answers in one message, the agent runs the whole batch on that answer, and the spend that follows is the spend somebody chose. Compute billed per use is spent the moment the request lands, so the pause has to come before the run, not after the invoice. That is the digital andon cord pointed at a budget instead of a defect.
Agent Capabilities
3 actionsInference
1- Run Model Runs any model in the Segmind catalogue by its slug and returns the generated output on the same request, with no polling. Image, video, and audio models return binary data that is stored in FlowRunner file storage and returned as a URL, while text models return their JSON directly. Every model takes a different input set, so pair this with Get Model Schema rather than guessing at fields.
Models
2- Get Model Schema Retrieves the published reference for a single model: its endpoint, category, whether it responds synchronously, its average cost, its average latency, and the full list of input fields with types, defaults, and allowed values. This is the operation that lets a flow price a run before committing to it.
- List Models Lists catalogue models with slug, description, modality, average cost, and average latency, optionally narrowed to one modality or filtered by a search term matched against name and description. Slugs returned here pass straight into Run Model.
Frequently Asked Questions
What can FlowRunner do with Segmind?
FlowRunner agents can run Run Model, Get Model Schema, and List Models in Segmind.
Does connecting Segmind to FlowRunner require OAuth?
No. Segmind connects to FlowRunner with an API key, no OAuth flow required.
Can Segmind trigger a FlowRunner workflow automatically?
Segmind doesn't currently expose triggers in FlowRunner. It connects as an action step inside workflows started by another trigger.
Start building with Segmind
$100 in credits. No card required. Connect in minutes.