A single API call is rarely the job. The real work is a sequence: authorize an upload, attach the file, then confirm the list. Kick off a migration, then poll until it finishes. Create a resource, then read it back to verify it stored. An OpenAPI spec describes the individual operations, but it says nothing about the order. Arazzo is the artifact that fills that gap — an OpenAPI-family specification for describing multi-step API workflows, where the output of one call feeds the input of the next.
That choreography matters for two audiences. Integrators get a tested, forkable recipe instead of stitching calls together from prose. Agents get an explicit plan they can execute step by step, with the data dependencies already declared.
Who publishes them
Arazzo is not a niche experiment in this catalog. Across APIs.io, 459 providers publish Arazzo workflows, totaling 5,104 individual workflow files. Coverage is broad and the leaders are deep:
- Stripe leads single-provider publishers with 31 workflows.
- Microsoft Azure API Management (29), Twilio (28), Affinda (27), and Mux (26) follow.
- Paystack, JFrog, Apigee, and Amazon Neptune each carry 25.
- Stytch, ServiceNow, and GitLab sit at 24 apiece.
There is also a cross-provider set of 255 workflows filed under multi — recipes that chain more than one provider together, such as classifying a message with Claude and opening a Zendesk ticket, or turning a Civitai image into a Notion entry. These are the workflows that treat the whole catalog as one composable surface.
What a good one looks like
The strong workflows are small, real, and honest about state. Look at GitLab’s set: “Start a Migration and Poll to Completion” kicks off a direct-transfer migration and then polls the import until it finishes — the polling loop is the whole point, and it is modeled explicitly. “Clean Up a Merged Branch” verifies a branch has been merged, then unprotects and deletes it, refusing to delete blindly. “Publish and Verify a Broadcast Message” creates a banner, then reads it back to confirm it was stored.
The pattern is consistent: each workflow names its steps, threads outputs into the next input, and includes a verification or guard step rather than assuming success. That is exactly what an agent needs to run a task unattended — and exactly what a human wants when forking a recipe into production.
Takeaway
Arazzo turns “here are the endpoints, good luck” into “here is the tested sequence.” With 5,104 workflows across 459 providers, plus 255 cross-provider recipes, the catalog is quietly building a library of executable integration knowledge. Browse it at apis.io/arazzo/ and fork a workflow instead of reverse-engineering one.