AI/ML API

Reach the AI/ML API catalog of more than 800 models through one FlowRunner connection. Agents run chat completions, image generation, and embeddings without a separate account per model provider.

5 actions API key available

AI/ML API website ↗ Platform Documentation ↗ Capability data verified 2026-08-01

The support knowledge base picks up a batch of newly published articles

Agent pulls the article text and strips boilerplate

Agent calls Create Embeddings to turn each article into a vector

Agent compares the embedding model ID against the model that produced the existing index

Agent posts the mismatch and the number of stored vectors that would be affected

The knowledge base owner decides whether to rebuild the whole index under the new model or hold on the old one

What This Integration Enables

AI/ML API is for teams whose bottleneck is not building the flow, it is getting permission to try a second model. The catalog runs past 800 models across chat, image, and embedding families, and List Models reads it live, so the roster your agent can choose from is the roster that exists today rather than the one that existed when someone wrote the flow. That turns model selection into something a flow can decide at runtime, using the same credential and the same billing relationship for every choice it makes.

The connector covers four distinct request shapes, which matters more than the raw count suggests. Chat completions handle the OpenAI-compatible majority. Create Anthropic Message exists because AI/ML API exposes Claude models through the native Anthropic Messages format rather than folding them into chat completions, so extended thinking and Anthropic-style tool use survive the trip. Generate Image and Create Embeddings cover the other two shapes. FlowRunner's BYOK model means the key stays yours: AI/ML API bills you directly, and FlowRunner never sits between you and your model spend.

Without FlowRunner

One account per model family A separate contract, key, and invoice for every provider a team wants to try

Model choice frozen at build time Switching to a newer model means reworking the flow that calls it

Evaluation is a project Trying a cheaper model against real traffic requires a procurement conversation first

With FlowRunner

One credential, one catalog Chat, image, and embedding models reachable through a single connection

Model as a runtime parameter The model ID is data the flow reads, not a rewrite the flow requires

Evaluation is a step A candidate model can run beside the incumbent inside the same flow

Use Case Scenarios

Running a challenger model beside the incumbent

A document classification flow has used the same model for months. Nobody wants to swap it without evidence. The agent sends every incoming document to the production model as usual, then sends a sampled share of the same documents to a challenger model chosen from List Models, and writes both classifications plus the disagreements to a table in Airtable. Nothing in the live path changes. After a few thousand documents, the operations lead has a real disagreement rate to look at rather than a vendor benchmark.

Retrieval over a support corpus

Create Embeddings converts published help articles into vectors that get written to a vector store such as Pinecone or Qdrant. When a ticket arrives in Zendesk, the agent embeds the ticket text with the same model, retrieves the nearest articles, and passes them to Create Chat Completion as grounding for a draft reply. The draft goes to an agent queue rather than to the customer, so a person still sends it.

Reasoning work routed to the Anthropic endpoint

A contract review flow needs a model that can work through a long document and explain its reasoning. The agent uses Create Anthropic Message for that step, because the native Messages format is where extended thinking lives, and uses the cheaper chat completion path for the routine extraction steps around it. Both run on one key. The expensive model is used where the difficulty actually is, not everywhere by default.

Human-in-Loop Highlight

The gate on this connector is the embedding index, not the chat output. Vectors produced by different embedding models are not comparable, so changing the model behind Create Embeddings silently invalidates every vector already stored. Retrieval does not error out when this happens. It just quietly gets worse, and the first person to notice is a customer reading a bad answer. So when the agent detects that the model ID configured for embedding no longer matches the one that produced the existing index, it stops before writing a single new vector and asks the knowledge base owner in Slack: "The embedding model changed from the one that built this index. Rebuilding means re-embedding every stored article and the partial state is not searchable. Rebuild now, revert to the previous model, or hold?" Rebuilding an index is expensive and cannot be half done. That is exactly the kind of decision an agent should hand back rather than make on momentum. This is the digital andon cord applied to a retrieval pipeline.

Agent processes routinely

Detects exception requiring judgment

Clear match Continues automatically

Ambiguous Routes to human via Slack

Human decides

Agent resumes with decision

Agent Capabilities

5 actions

Language Models

Create Chat Completion Sends a prompt or a full messages array to any OpenAI-compatible model in the catalog and returns one complete response. Supports tool calling and JSON Schema structured output, which is what makes the result usable as data by the next step in a flow rather than as text a human has to read.
Create Anthropic Message Calls a Claude model through the native Anthropic Messages format that AI/ML API exposes separately from chat completions. Used when a step needs extended thinking or Anthropic-style tool use rather than a plain completion. Max Tokens is required by this endpoint.

Images

Generate Image Generates an image from a text prompt using any image model in the catalog, including GPT Image, Imagen, Seedream, FLUX, and Nano Banana. Model and Prompt are the only fields every image model shares, so per-model controls travel through Extra Options. Used for on-demand creative inside a flow that already knows the brief.

Retrieval

Create Embeddings Converts a batch of text inputs into vectors in one call, returning one vector per input in the order supplied. Used to build and maintain the semantic index behind ticket deflection, document routing, and duplicate detection.

Catalog

List Models Reads the live model catalog with each model's display name, developer, context length, and the endpoint family that serves it. Filter by capability to narrow to chat, Claude, image, or embedding models. Used when a flow selects a model at runtime instead of hard-coding one.

Frequently Asked Questions

What can FlowRunner do with AI/ML API?

FlowRunner agents can run Create Chat Completion, Create Anthropic Message, and Generate Image in AI/ML API, plus 2 more actions.

Does connecting AI/ML API to FlowRunner require OAuth?

No. AI/ML API connects to FlowRunner with an API key, no OAuth flow required.

Can AI/ML API trigger a FlowRunner workflow automatically?

AI/ML API doesn't currently expose triggers in FlowRunner. It connects as an action step inside workflows started by another trigger.

Start building with AI/ML API

$100 in credits. No card required. Connect in minutes.

Start Building Free Talk to us