Skip to main content

App Studio

Model serving

App Studio hosts the things people open rather than call: a Streamlit dashboard over an agent, a Gradio demo, a FastAPI service, a container you built yourself. Point it at a repository, pick a framework and a size, and Studio runs it and gives it an address on the platform.

Who can use it: the Builder role or higher, in a tenant licensed for model serving. Without that licence the sidebar shows App Studio locked, and the address refuses to load rather than failing halfway.

What an app is here​

An app is your code, run as a pod on the cluster and served at a path on the platform's own host. It is not an agent, and Studio does not turn one into the other. The two meet in one place only: an app is promoted between environments through the same approval machinery an agent uses.

There is no image build step yet. Studio clones the repository into the pod and installs the dependencies when the pod starts, which is why a first boot takes longer than a restart and why a dependency error appears as a build failure rather than a code error.

The list​

App Studio: the summary line and New app, the status counts All, Running, Building, Degraded, Failed and Stopped, the search box and the All frameworks filter, and the apps table with each app's address and framework, status, traffic, its deploy as a pinned commit, branch and owner, and resources, with Open on each running app.App Studio: the summary line and New app, the status counts All, Running, Building, Degraded, Failed and Stopped, the search box and the All frameworks filter, and the apps table with each app's address and framework, status, traffic, its deploy as a pinned commit, branch and owner, and resources, with Open on each running app.
App Studio lists the apps deployed from repositories and images, with their status, traffic, deploy and compute. New app starts the wizard.

The header counts apps by status, and each count is a filter: All, Running, Building, Degraded, Failed, Stopped. A framework filter appears only when the workspace has more than one framework in it. The table's columns are the app, its status, traffic, the deploy and its resources.

Open appears only on a running app. The row menu carries View details, Settings, Start, Stop and Delete. An app that is mid-build offers neither Start nor Stop, because neither is a safe thing to do to it. Delete is disabled on an app someone else owns, and names the owner, because Studio scopes every change to the caller and the delete could never have succeeded.

The list refreshes every 5 seconds while any app is running, starting or building, and stops once every app is stopped or failed. It pauses while the browser tab is hidden and resumes on its next refresh after you return, not at once. Two empty results in a row are required before it will show you an empty table, so a build that briefly reports nothing does not blank the page.

Two different empty results read differently on purpose. No apps match these filters means your filters excluded everything. A service error says so in as many words, and adds that your apps, if you have any, are not being shown.

Before there is anything​

The App Studio first-run page with a Deploy your first app card and a New app button, a row of framework tiles for Streamlit, Gradio, FastAPI, Dash, Panel, Flask, Node/React and Docker, two path cards for a Git repo and a ZIP upload marked coming soon, and a Before you ship list covering compute, registry, routing and budget.The App Studio first-run page with a Deploy your first app card and a New app button, a row of framework tiles for Streamlit, Gradio, FastAPI, Dash, Panel, Flask, Node/React and Docker, two path cards for a Git repo and a ZIP upload marked coming soon, and a Before you ship list covering compute, registry, routing and budget.
A workspace with no apps yet: the eight frameworks you can start from, the two ways to ship, and what the platform already has in place.

A workspace with no apps gets the eight frameworks as a starting point, the two ways to ship, and a short list of what the platform already provides. Read that list as a description of the platform rather than a check of your cluster: nothing behind it is tested live. Set up beside the budget line is the one row that does anything.

The Git card's hint reads ~2 min · auto-rebuild on push coming soon. Automatic rebuild on push is not built, and the card says so, as the wizard's own source cards do.

The statuses​

The list and an app's own page read the same status, in the same words. They used to keep separate vocabularies, so one state had two names.

StatusWhat it means
BuildingThe pod is being created and dependencies installed
RunningThe pod is up and answering
DegradedServing, but not on every replica, so check the logs
FailedThe build or the boot did not finish
StoppedDeliberately scaled to zero, and startable again

A state the service never writes falls back to Idle, which is why there is no filter for it: the chips cover exactly the statuses the service can write.

The address​

An app is served at the platform's own host under a path, as https://<your platform host>/apps/<name>/. It is not a subdomain, and a subdomain will not resolve. The name gets a short unique suffix, so two people can use the same name.

The address is deliberately on the same origin as Studio, which lets the Preview tab embed the app. The same-origin choice would otherwise hand the embedded page the camera, microphone, screen capture and location permissions that a page grants itself, so the platform denies all four to every app it embeds and allows only the clipboard and full screen.

An app's address does not contain the owner's email. An older build put it there, which leaked it into every shared link.

What to do next​

Describes the platform as of .

Was this page helpful?