A Claude artifact made today can keep 100 MB of data, know who is looking at it, and call that person's connectors. It still has no backend. Both facts are true, and the second decides more than the first.
"Artifacts are only for prototypes" was a fair rule until 16 September 2026. Anthropic now calls anything made in chat before that date a legacy artifact, and the limits most people remember (20 MB of text, storage only after publishing, data deleted on unpublish) belong to that older kind.
The replacement rule is narrower. An artifact is a good permanent home for an interface whose users all sit inside one Claude organisation and whose data lives somewhere else. It is the wrong home for a system of record, for anything that must run while nobody has the page open, and for anything a customer without a Claude seat has to use.
Prototype versus production is the wrong axis. Ownership is the right one.
Each row is sourced to Anthropic's Help Center or Claude Code docs. The full reference is the Claude artifacts reference leaf.
| Capability | What the docs say | Source |
|---|---|---|
| Data storage | Personal or shared. 100 MB and 25,000 documents per artifact, whichever comes first | Help Center, artifacts article |
| Uploaded files | 20 MB per file (2 MB for SVG), 5,000 files and 1 GB per artifact | Help Center, artifacts article |
| Connectors | Reads and writes through the viewer's own connected apps. Pro and above. First use needs approval | Help Center, artifacts article |
| Calls to Claude | Count against each viewer's own plan limits | Help Center, artifacts article |
| Sharing | Depends on plan and admin settings. Up to 50 outside invitees per artifact | Help Center, sharing article |
| Governance | Retention policy; audit events under claude_artifact_*; three Compliance API endpoints (Claude Code surface) | Claude Code docs |
| Legacy artifacts | 20 MB, text only, storage only after publishing; unpublishing deletes the data | Help Center, artifacts article |
That is a real application surface. For an internal tracker, a review queue or an ops dashboard, standing up auth, hosting, a deploy pipeline and an inference bill to replace it is cost with no return.
Anthropic's Claude Code docs describe an artifact as one self-contained page that cannot serve multiple routes or authenticate viewers itself. For a hosted internal tool with a backend, the docs point to your own infrastructure. Nothing in the docs describes scheduled jobs, webhooks or server-held secrets.
Per the same docs, network calls reach only the page's own origin, plus Google Fonts. Scripts load from five public CDN hosts. Any other outside call has to go through a connector.
Connector calls use the viewing account's connection, so two people can open the same page and see different data. Connector tools that need approval for each action are not available to artifacts at all.
An artifact that connects to other websites cannot be shared by public link or email invitation. Claude and connectors do not work for outside link viewers.
Chat artifacts became legacy on 16 September 2026. Standalone Claude Design closes on 14 December 2026. Claude Code cannot publish artifacts from Amazon Bedrock, Google Cloud or Microsoft Foundry sessions.
| Question | Stay on artifacts | Migrate |
|---|---|---|
| Who are the users? | All inside one Claude organisation | Customers, the public, or anyone without a seat |
| Where does the truth live? | In an ERP, CRM or database the page reads through a connector | In the artifact's own storage |
| Does anything run when the page is closed? | No | Schedules, webhooks, queues, background agents |
Three "stay" answers and the artifact is not a prototype. It is the cheapest correct architecture. One "migrate" answer and it is a prototype, whatever it looks like.
The trade-off on the "stay" side is real: you give up tests, environments, CI and an SLA, and you inherit Anthropic's roadmap. That is acceptable for a view you can regenerate in an afternoon. It is not acceptable for the ledger.
The durable shape is an artifact as a disposable interface over MCP servers that front the real systems. Permissions, validation and writes live in the MCP server and the system behind it. The page holds layout and nothing else.
This is the harness model applied to interfaces. One operator regenerates the page when the workflow changes. The human specialists and the audit trail sit behind the connector, where they belong.
Test it with one question: if Anthropic retired this artifact type tomorrow, what would you lose? If the answer is "a layout", you built it right.
Move records out of artifact storage into a database you control: D1 for edge-local SQLite, Postgres for anything relational and shared.
Move rules, writes and integrations into a Worker, and expose them as an MCP server so the existing artifact keeps working against the new backend.
Put your own auth in front of the backend the day a non-Claude user needs access.
Rebuild the page on Pages only when one of the three questions flips. Many never do.
Steps 1 and 2 remove the risk. Step 4 is optional, and skipping it is often the right call.
Residency. Anthropic's Claude Code docs state that artifact content is stored on Anthropic-operated infrastructure. The Team and Enterprise admin guide gives no residency commitment for artifacts and says Claude Design does not support residency requirements. Treat personal information in artifact storage as a POPIA Section 72 cross-border transfer, the same position this tree takes on Claude inference.
Cost. Artifact usage draws on each viewer's seat, or on API rates for usage-based Enterprise. There is no hosting line item, which makes an internal tool look free in rand terms. The cost has moved to seat limits, and a heavy shared dashboard will show up there.
Regulated clients. The interface can be an artifact; the records cannot, wherever a South African client has sector rules on where records live.
Checked 21 November 2026. Verify the limits against these pages before you commit to a migration.