What an Artifact gives you
On Team and Enterprise plans, a person can publish what they built in Claude to a page and share it inside the company. Anthropic’s admin guide describes it as “a shareable Anthropic-hosted page”.
For a company that already pays for Claude, that covers a lot:
- Sharing stays inside the organization unless an admin allows external links.
- Access follows your identity provider. Removing someone from the organization removes their access to its artifacts.
- Admins have one switch. Artifacts can be turned off for the whole organization, and existing links stop working.
- Pages can read connected apps. With artifact connectors on, an artifact can read from and write to apps a person has connected.
- There is nothing to deploy. No server, no DNS record, no ticket.
If a tool fits inside that, an Artifact is the simplest sanctioned place for it. Use it.
Four signs an app has outgrown one
1. It needs its own backend
An Artifact is a page. Some apps are a page plus a service: an API that several screens call, a job that runs on a schedule, a queue, a process that keeps state between requests. The admin guide describes pages, templates and connectors. It does not describe a way to run your own server process next to the page.
2. It is a long-running Python server
Streamlit, Dash and similar frameworks are common output from Claude Code when the request is “make me a dashboard”. They run as a server process that stays up and holds a session for each viewer. That is a different shape from a published page.
3. Its data access should not belong to one person
The admin guide is specific about how a shared artifact reaches data: “Each user connects their own apps, even in a shared artifact.” The exception is a live-shared dashboard, which “shows results from its owner’s connection.”
Both models tie data access to a person. Either every viewer needs their own access to the source, or everyone sees the data through the owner’s access. Many internal apps need a third arrangement: a read-only credential that IT registers once, that belongs to no individual, and that the builder never sees.
4. It has to run in your cloud
Artifacts are hosted by Anthropic. The admin guide also says they “aren’t available through third-party cloud platforms.” If your security policy says an app that reads production data must run in an account you control, an Artifact cannot meet that policy, however well it is governed.
A short test
Ask these about the app. One “yes” is enough.
- Does it run code on a server, apart from what runs in the viewer’s browser?
- Does it need to stay running between visits?
- Should it read production data through a credential that belongs to no individual?
- Does policy require it to run in your own cloud account?
Four “no” answers mean the app belongs in an Artifact.
Where an app that has outgrown one can go
Each of these is in use today.
| Option | Good fit when | What you take on |
|---|---|---|
| Your own pipeline | You have a platform team with time | DNS, certificates, single sign-on and secrets for every app |
| A general host such as Vercel | The builders are engineers and the app is a web front end | Access rules and data credentials are yours to arrange. See Vercel for internal apps. |
| Your data platform’s app hosting | All the data is on that one platform | Apps are tied to the platform. See Databricks Apps. |
| A governed app platform | You want to bring apps into a vendor’s platform | The app lives in that platform. See Superblocks. |
intrium is being built for the same gap: the app runs as written at a URL under your own domain, opens only to an Okta or Google group, and reads Postgres or Snowflake through a credential the builder never sees. It is pre-launch. intrium is not affiliated with Anthropic.
One caution on the fourth sign. intrium is hosted by us to start, and running in your own AWS account is planned as the Enterprise option. If own-account hosting is a hard requirement today, your own pipeline or a platform that deploys inside your network is the answer for now.
What to do before choosing
Count first. Ask the people who own your Claude Code rollout how many apps staff have built, and put each one through the test above. If nearly all of them pass as Artifacts, you may not need anything else. If a handful do not, those are the ones that will end up on a personal account or a public link unless there is an approved place for them. A written rule helps here, and there is a policy template you can adapt.