Skip to main content
The Prompt Hub is the centralized repository for every Prompt Object your team owns. It’s the source of truth — the place every other Arize AX surface (Playground, experiments, the SDK, your application at runtime) reaches into when it needs a prompt. If the Prompt Object is the unit and versioning is the control plane, the Hub is the storage layer that holds them all and makes them addressable.
Arize AX Prompt Hub showing a list of prompts including trip-planner variants and existing prompts, with columns for Name, Description, Tags, Versions, Created By, Last Updated, and an Optimize action button per row

What lives in the Hub

The Hub stores Prompt Objects, not raw strings. Each prompt in the Hub has:
  • A name (e.g., support-agent, tool-selection-router).
  • A description that captures intent (“Routes inbound queries to the right downstream skill”).
  • A version history — every save lands here as a new immutable record.
  • Tags that point at specific versions (production, staging, dev, prod-v1.2, …).
  • The current state of all five Prompt Object parts at each version.
You can search and filter the Hub by name, tag, model, and last-updated time. Tags make it possible to ask the Hub things like “show me every prompt that has a production tag and uses gpt-5.4.”

Five pillars of the Hub

The Hub is shaped around five guarantees:
  • Unified prompt ecosystem — stop hunting through Slack messages or codebase comments. Every Prompt Object — template, parameters, tools, response format — lives in one searchable place. Past, present, and future versions sit together.
  • Immutable version control — every save creates a hashed, immutable version. Tags move; versions don’t. Rollback is a tag move, not a restore.
  • Rapid experimentation and comparison — load any version into the Playground, run against a dataset, score with evaluators, compare side-by-side, save the winner back as a new version. The loop closes inside one tool.
  • Compatible, easy-to-use SDK — Python, TypeScript, Go, and CLI clients can list, fetch, create, and tag prompts. Your application reads from the Hub at runtime; your CI script can push new versions and move tags.
  • Path to automated optimization — once versions and eval feedback live in the Hub, Prompt Learning and other optimization paths can consume them as inputs.

Where the Hub sits in the platform

Three other Arize AX surfaces talk to the Hub:
SurfaceWhat it does with the Hub
Prompt PlaygroundLoads any prompt version for interactive iteration; saves new versions back when you save a Playground state.
ExperimentsAn experiment run records which prompt version (by hash) was tested. That hash is permanent — re-running the experiment always uses the same input.
Your applicationReads prompts via the SDK by name + tag (or by name + version hash). The SDK fetches from the Hub and caches locally.
The Hub’s role is to be the one source-of-truth that all of these reach. If two teammates load support-agent with tag production, they get the same version. If you re-run an experiment from three weeks ago, the same prompt version comes back. If you move the production tag, your live application picks up the new version on its next refresh.

What the Hub is not

A few things the Hub doesn’t try to be:
  • Not a prompt engineering tool. The Playground is where iteration happens. The Hub is where results land.
  • Not a runtime execution surface. The Hub stores Prompt Objects; it doesn’t run them. The Playground, experiments, and your application do the running.
  • Not a folder of strings. Don’t think of it as a file tree of template text. Think of it as a database of versioned objects with rich metadata.

Reaching the Hub

You can interact with the Hub from anywhere that suits the moment:
  • UI — browse, search, diff, tag, and edit visually from the Prompts section in the left nav.
  • Playground — load and save during iteration.
  • SDK — fetch in your application, push new versions from your CI script.
  • CLIax prompts list, ax prompts get, ax prompts create-version, etc.
  • REST API — same operations available over HTTP.
The next page covers the SDK path in depth, including the local-cache pattern that keeps the Hub off your inference hot path.

Next step

The Hub is where prompts live. The next question is how your application reaches them safely.

Next: Loading Prompts in Your Application