Live Client Dashboards from a Prompt with Claude Artifacts

Live client dashboards from a prompt stopped being a stretch goal for Claude Artifacts in mid-July 2026. In the update window of July 13–17, Anthropic shipped the missing piece: a published Claude Code Artifact can now call MCP connectors every time someone views it, so the page shows live data — and can even take actions — instead of a frozen snapshot from the session that built it.

For agencies, this lands in a genuinely awkward gap in the tooling market: the space between a static PDF-ish report a client skims once and a full BI-tool license the agency pays for every month. The same update window also brought public sharing links, editor roles and multiplayer editing — which sounds like the whole client-portal problem solved in one release.

It is not, and the reason is the single most under-covered fact in all the coverage we surveyed: per Anthropic's own documentation, an artifact that calls connectors cannot be shared to a public link on any plan. Live data and no-login client access are mutually exclusive on the same page. This guide covers what actually shipped, how viewer-side connector calls work, what an artifact still is not, and a practical decision framework for when a prompt-built dashboard beats a real BI tool — and when it does not.

Claude Code Artifacts launched in beta on June 18, 2026 — we covered the launch in our June guide to shipping a coding session as a shareable live web page. At that point an artifact was org-private or nothing, and every page was a static capture: whatever data the session gathered while building, frozen at publish time. We flagged both as hard limits then. Twenty-five days later, the July 13–17 update window delivered both — through Claude Code releases v2.1.207 to v2.1.212, per Anthropic's Week 29 digest.

Independent coverage from AlternativeTo and AlphaSignal corroborates the July 13 ship date and the framing: public artifact sharing plus real-time collaboration. But the three features are separately gated, and the gating is where agency plans meet reality.

One disambiguation before going deeper: this post is about Claude Code Artifacts — pages published from coding sessions — not the 2024-era Artifacts feature inside the claude.ai chat interface, which is a separate system with its own persistent-storage mechanism. The two share a brand name, not a feature set, and conflating them is the fastest way to promise a client something the tool cannot do.

The design decision that makes this feature safe is also the one that reshapes how you deliver it: connector calls from a published artifact run through the viewing account's MCP connections — not the publisher's. Per Anthropic's documentation, two people opening the same dashboard can see different data depending on what their own accounts can access.

The flow, as documented: before a page's first connector call, the viewer must approve connector access. A viewer who declines — or who simply has not connected the needed connector — still sees the page, just without its live sections. The page itself never sees anyone's credentials; claude.ai makes the connector calls on the page's behalf. Responses are cached in the viewer's browser, so a reopened dashboard renders instantly from cache and then updates with fresh results.

The same rule extends to write actions. A dashboard button that posts a Slack message or updates an issue runs that action through the viewer's account, not the publisher's. For agency work this is a genuine accountability feature — the action is attributed to the person who clicked, with their permissions — but it also means the client must have the relevant connector attached to their own claude.ai account before the live sections light up. Onboarding the client's connectors becomes part of the deliverable.

Here is the fact that every secondary write-up mentions in passing and none foregrounds: the two marquee features of this update cannot coexist on the same artifact. Anthropic's documentation is unambiguous.

Play that out against the plan matrix and the practical shape emerges. On Team and Enterprise, a connector-backed dashboard can be kept private or shared inside the organization — which works when your client's stakeholders sit inside the same claude.ai org, and does not when they are external. On Pro and Max, where a public link is the only sharing mechanism that exists, a connector-backed artifact simply stays private to its author. A solo consultant on a Pro plan literally cannot share a live-data dashboard with anyone.

The consequence for client delivery: a live-refreshing dashboard for a client with no claude.ai login is not a thing Claude Artifacts can produce today. Your options are (a) a public, no-login artifact that is a static snapshot — refreshed by republishing, (b) a live connector-backed artifact for viewers inside a shared org, or (c) a real reporting or BI tool. That three-way fork is the backbone of the decision matrix in section 06.

Anthropic's own framing of what an artifact is has not moved since the June beta, and it is worth quoting the documentation's language precisely: an artifact is "a capture of work, not an application." It is one self-contained page with no backend, so it cannot store form input or serve multiple routes — and its only path to outside data at view time is calling MCP connectors.

The page constraints are likewise unchanged. A strict Content Security Policy blocks external scripts, stylesheets, fonts, images, fetch, XHR and WebSocket calls to any other host — CSS and JavaScript are inlined and images embedded as data URIs. The sole exception is connector calls, which the page hands to claude.ai rather than making itself. Source files are .html, .htm or .md; relative links do not resolve because nothing else is deployed alongside the page; and the rendered page is capped at 16 MiB, with oversized embedded images the usual cause of a failed publish.

Two practical notes complete the picture. First, availability is a checklist, not a default: a Pro/Max/Team/Enterprise plan, a claude.ai-account-backed session, the Anthropic API as model provider (artifacts are not available via Amazon Bedrock, Google Cloud's Agent Platform, or Microsoft Foundry), and no CMEK/HIPAA/Zero-Data-Retention org policy. Second, building a page costs output tokens like any other response, and a styled interactive page is more token-intensive than the same content as terminal text — inline CSS, interactive JavaScript and embedded images are the drivers. A dashboard you regenerate daily has a real, if small, unit cost.

Filter the feature set through the constraints and a concrete service menu falls out. The strongest fits share two traits: the data lives in tools the viewer can connect via MCP, and the audience either accepts a snapshot or sits inside your org.

• Kickoff and audit reports as interactive pages. A one-off deliverable — a technical audit, a campaign retrospective, a migration plan — published as a designed, interactive page instead of a PDF. No live data needed, so it can be a public link the client opens with zero login. Republish to refresh.
• Internal ops dashboards over the org's own tools. Sprint status from the connected repo, support-queue health, content-calendar coverage — live on every view via connectors, shared inside the org. This is where the viewer-side model shines: each teammate sees what their access allows.
• Client dashboards inside a shared workspace. For retained clients on a shared Team/Enterprise org, a connector-backed page over their own systems becomes a standing live surface — with write actions (post an update, file an issue) running as the person who clicks.
• Slack-native reporting via Claude Tag. With Claude Tag and artifacts enabled org-wide, asking for a dashboard in a Slack thread returns a working, org-scoped page — reporting that starts where the conversation already is.

Two force multipliers are easy to miss. Anthropic ships a built-in design skill (CLI v2.1.182+) that gives pages a deliberate palette, typography and layout without extra prompting — and it checks the project's CLAUDE.md or a repo theme file for existing design tokens first, which means a client dashboard can inherit brand colors automatically. And a dashboard prompt you reuse weekly is a candidate for packaging into a skill — our guide to training Claude agents by demonstration covers that pattern. Pair it with webhook-triggered managed agents and the weekly republish of a snapshot dashboard stops being a manual chore.

This is exactly the shape of work we build in our CRM & automation engagements: reporting surfaces over the systems a client already runs, delivered as working software in days rather than a BI implementation project in quarters.

No coverage we surveyed frames this as what it is for an agency operator: a buy/build/prompt decision keyed to client count, data sources and access model. The matrix below is our house framing — capability facts from Anthropic's documentation, tool pricing as reported by 2026 competitor analyses (ReportingNinja, 1ClickReport) and Databox's 2026 BI roundup, and best-fit judgments from our own delivery experience. Pricing cells are directional, not quotes — verify against each vendor before budgeting.

Sources: Anthropic Artifacts documentation (capabilities and limits, retrieved July 24, 2026); AgencyAnalytics pricing as reported by ReportingNinja and 1ClickReport 2026 analyses — figures are reseller-analysis-stated, not vendor-audited; Databox's 2026 BI alternatives roundup for the connector-cost range; best-fit column is Digital Applied's house judgment.

The pattern in the matrix is not subtle. Claude Artifacts wins wherever the data already lives in MCP-connectable tools and the audience is internal or snapshot-tolerant — because the marginal cost of a dashboard collapses to tokens. BI and reporting tools win wherever you need scheduled refresh into a public surface, ad platform data pipelines, or one pane across dozens of clients. Notably absent from every vendor comparison we found: an Artifacts row at all. The category is moving faster than the shootout posts.

Anthropic shipped this with an unusually complete admin surface for a month-old capability, and for agencies working inside client orgs — or running their own Team plan — the controls matter as much as the features.

• Layered toggles. Artifacts overall switch on and off under Settings > Claude Code > Capabilities; connector calls from artifacts have a separate "Enable artifact connectors" toggle under Settings > Capabilities. You can allow published pages while keeping live connector calls off.
• Plan-level defaults. Team plans have artifacts on by default; Enterprise requires an Owner to enable them, and Enterprise with RBAC can scope the Artifacts permission per role.
• Retention and audit. Admins can set separate data-retention periods for private versus shared artifacts, and every publish, share and delete event logs under claude_artifact_* audit-log event types.
• Compliance API. An org can programmatically list its artifacts, fetch a specific version's content, and delete an artifact — the hooks a compliance team needs before green-lighting client-facing use.
• Network allowlisting. Viewers load pages from a sandboxed *.claudeusercontent.com origin, separate from claude.ai — orgs restricting outbound access need both domains allowlisted.

The forward-looking read: this governance surface is what enterprise-readiness looks like when a vendor intends a feature to carry real workloads. Anthropic went from beta to live external data, multiplayer editing and a compliance API in under a month. The sensible planning assumption is that the current hard edges — single page, no backend, the public-link exclusion — are the next things to move, and an agency that builds the delivery muscle now inherits each loosened constraint for free.

Collapse everything above into four calls an agency operator has to make. The variable that decides each one is almost never "which tool makes prettier charts" — it is who the viewer is and where the data lives.

Our interpretation of the trend: viewer-side connector calls are the first mainstream implementation of a pattern we expect to define agency reporting over the next few years — dashboards as generated, disposable surfaces over live systems, rather than as standing infrastructure someone maintains. When a dashboard costs tokens instead of a license and a build sprint, the economics of "just make a page for this meeting" invert. If you are working out what that means for your own client-reporting stack — or where agentic tooling genuinely replaces SaaS line items and where it does not — that scoping conversation is the front door of our AI transformation engagements.