Step 5: Build the Next
Goal: Describe an outcome; let any AI propose the whole build - including a real feed.
Step 5 of 5. The payoff. The same MCP server any AI reads from is the one it builds with - and not just warm queries. An AI can propose new data connections, API sources (real third-party feeds), surfaces, and the wiring between them. You describe the outcome; it proposes the shape; you accept.
Ask for an outcome, not a step
From the agent you connected in Step 4 (or Azi), ask the way you'd say it out loud:
"We want to watch community growth next to signups. Add a live feed of our GitHub repo's stars, combine it with product signups, and show me both over the last day."
You did not name a node type or write a query. The agent figures out the shape:
- It explores what exists (
list_data_connections,get_workspace_flow). - It proposes a new API Source pointed at the public GitHub endpoint (it polls on a schedule -
no credentials, it is public), wires it into your
GrowthMetricssurface, and proposes a new warm query that combines stars with signups - all in one set of proposals (manage_proposals), testing the query first withrun_adhoc_warm_query.
You accept; then real data flows
The proposals land in your workspace, tagged as AI-made. You accept them in the portal and deploy (accept always stays human - that is the boundary). The API Source starts pulling real data on its schedule - a live feed, no simulator, no credentials to wire. One natural-language ask got you a new real source, the wiring, and a combined source of truth.
(Want a custom feed instead of a public API? The AI can also create a data connection and push data into it from its own process - same idea, different source.)
Watch it compound
A simulated feed and a real public feed, combined into one source of truth - proposed by an AI, accepted by you, reachable by every AI on your team. That is the whole loop: connect, crystallize, build - by hand, with Azi, or with any agent you bring, and you hold the proposals. Keep going and the workshop compounds.
Where to go next
- Go deeper on each workbench -> Guides
- Build and share your own workbench -> Host Your Own Workbench