Doesn't Anthropic plan to solve this?

Anthropic's sharing model is designed around Claude as a product. Their Team and Enterprise tiers allow org-internal sharing — but viewers need Claude seats. The Publish button creates public URLs. There's no "only @yourcompany.com" middle ground. Building an identity-aware, multi-IdP publishing layer for arbitrary HTML is a different product category from an AI assistant.

Can't teams just use Confluence?

Confluence is a documentation editor. It has an authoring model (wiki-style editing, page trees, collaborative writing). It cannot host arbitrary HTML. Pasting a Claude Code artifact into Confluence produces a broken skeleton that requires 30–45 minutes of manual reformatting to approximate the original.

Is this problem specific to Claude Code?

No. Cursor, Windsurf, Codex, and every other AI coding agent writes HTML to disk. All of them have the same gap: no path from local HTML file to company-authenticated URL. Claude Code is the highest-volume source today (73% of daily AI tool users), but the problem exists across every tool in the category.

AI Agents Are Creating an Internal Publishing Problem Nobody Has Solved

What changed in 18 months

Claude Code reached $2.5B ARR and became the number-one AI coding tool in eight months. 73% of developers now use AI tools daily (Developer Survey 2026). 95% use them weekly.

These aren't vanity metrics. They describe a structural shift in how software engineering work gets done. AI tools aren't writing code alongside engineers anymore — they're producing the primary output of engineering sessions.

That output is HTML.

What AI agents produce

The default output of an AI coding session isn't a file to edit or a diff to review. It's a rendered experience.

Ask Claude Code to document an architecture decision and you get an interactive diagram with clickable components, animated dependency flows, and hover tooltips. Ask it to summarize competitive research and you get a sortable comparison table with embedded charts. Ask it to generate a sprint summary and you get a dashboard with tabs, filters, and responsive layout.

Self-contained. Single HTML file. All CSS and JavaScript inlined. Beautiful.

Completely impossible to share with your team.

Where they go to die

The artifact workflow for the majority of engineering teams today:

Generate the artifact in Claude Code
Open it locally to verify it looks good
Try to share it
Take a screenshot
Post the screenshot to Slack

Step 4 destroys everything that made step 2 impressive. The interactive architecture diagram becomes a flat JPEG. The sortable competitive analysis becomes a static image. The animated sprint dashboard becomes a frozen snapshot.

Or: don't screenshot. Give up. Leave the artifact on your local machine. The work product stays in the conversation, accessible to no one.

The screenshot is the worst outcome because it looks like you shared something. The content arrived in Slack. The team saw it. But the interactivity — the responsive layout, the live filters, the animated transitions — is gone. A Rembrandt as a Polaroid.

Why existing tools don't work

Confluence and SharePoint strip JavaScript and CSS. Pasting an HTML artifact produces a broken skeleton of the original. These tools are document editors. They cannot render applications.

Google Drive downloads the file. The viewer's browser opens a local copy with a file:// URL. Relative paths break. JavaScript security restrictions block network calls. The artifact doesn't render correctly.

GitHub Pages works for engineering audiences with GitHub accounts. Your PM, VP of Engineering, legal team, and sales team don't have GitHub accounts. Private repos require GitHub Enterprise Cloud — $21/user/month. For 100 viewers, that's $2,100/month.

Claude's Publish button creates a fully public URL on claude.ai. No company restriction. Anything sensitive — competitive analyses, financial models, architecture proposals, client data — cannot go to a publicly accessible URL, even an obscure one.

The viewer asymmetry problem

One engineer creates an architecture proposal. Seven people need to review it before Monday: the VP of Engineering, four product managers, the security lead, and the designer.

Exactly zero of these people should need to create a new account, install an app, or file an IT ticket to see a document their colleague created.

Per-seat pricing tools penalize this pattern. A tool that charges $7/viewer creates a $49/month incremental cost for seven new viewers — incentivizing the engineer to not share, or to share something degraded (the screenshot) that doesn't trigger the fee.

Flat pricing inverts this: seven viewers cost the same as one. The natural impulse to share reaches its audience.

What the gap looks like at scale

25% of new Linear issues are already agent-authored, growing 5x in three months. That's task-tracking. The same shift is happening in content production.

Every company running agent workflows for research, reporting, and documentation is generating HTML output that has no distribution layer. The volume grows with every AI tool adoption. The infrastructure to handle it doesn't exist.

The companies building agent workforces — where autonomous agents produce weekly competitive reports, generate sprint dashboards, create onboarding documentation — are creating a publishing backlog measured in hundreds of artifacts per week. Screenshots can't scale.

What the solution looks like

The publishing layer for AI-generated content needs three properties:

Programmatic. A function call or CLI command, not a manual upload workflow. Agents can't click "publish." They need an API.

Company-restricted. A URL that only people with the company's email domain can access. Not public. Not per-seat.

Flat-priced. A model where the 51st viewer costs nothing extra. Sharing should be economically frictionless.

That's gated publishing. One command, company SSO, permanent URL. The artifact gets to the people who need it.