5 ways to share what your AI built (and which is best)

AI tools have gotten remarkably good at building things. Claude produces interactive dashboards. ChatGPT's canvas generates polished reports. Codex writes entire single-page apps. The output is real, functional HTML — and then you hit the same wall every time: how do you get this in front of someone who wasn't in the chat?

There's no single right answer. The best method depends on what you built, who needs to see it, and how long the link needs to last. This guide covers all five approaches people use in 2026, with an honest assessment of where each one shines and where it breaks.

1. Screenshot or PDF export

How it works

The simplest possible approach: take a screenshot, export to PDF, or copy the text and paste it into a document. No hosting, no URLs, no setup.

What's good

• Instant. No tools, no accounts, no waiting. Command-Shift-4 and you're done.
• Universal. Everyone can open a screenshot. No browser compatibility, no rendering issues, no "can you see this?" debugging.
• Offline-friendly. The file exists on the recipient's device. No link to go stale, no server to go down.

What breaks

• Interactivity dies. That interactive chart becomes a picture of a chart. The calculator becomes a picture of a calculator. Hover states, animations, filtering — gone. You're sharing a corpse of the thing your AI built, not the thing itself.
• Updates mean re-sending. Change one number and you need to screenshot again, re-attach, re-send. The recipient is now managing file versions in their downloads folder.
• Fidelity loss. Screenshots compress, crop awkwardly, and miss content below the fold. PDFs flatten responsive layouts into a single fixed width.

2. Claude's native publish (or ChatGPT's share link)

How it works

Claude and ChatGPT both offer built-in sharing. In Claude, you click "Publish" on an artifact and get a claude.ai URL. In ChatGPT, you share a canvas and get a link on OpenAI's domain. The recipient opens it in a browser and sees the live, interactive artifact.

What's good

• One click. No setup, no external tools. It's built into the conversation interface.
• Interactivity preserved. The recipient sees the real thing — charts animate, buttons work, layouts reflow.
• Free. No additional cost beyond your existing AI subscription.

What breaks

• Each publish is a new URL. Update your artifact and publish again? New link. The old one is frozen. Recipients with the old bookmark see a stale version and don't know it.
• No access control. The link is public. Anyone with the URL can open it. For a fun demo, fine. For a proposal with pricing, a compliance report, or anything involving client data — that's a problem.
• It's branded as AI output. The URL lives on claude.ai or OpenAI's domain. That's transparent about the origin, but it frames your work as "something a chatbot made" rather than "something you delivered."
• No version history. No way to see what was live at a given date, no rollback, no audit trail.

3. Copy the HTML and self-host

How it works

Copy the artifact's source code, drop it into a GitHub repo (or any static host — Netlify, Vercel, Cloudflare Pages, an S3 bucket), and deploy. You now have a URL on your own domain that serves the artifact exactly as the AI built it.

What's good

• Full control. Your domain, your infrastructure, your deploy pipeline. You decide everything about how the content is served, cached, and secured.
• No vendor lock-in. The HTML is yours. The hosting is commoditized. Move it anywhere, anytime.
• Professional URLs. The link lives on your own domain. It looks like part of your website, because it is.

What breaks

• Requires developer skills. Git, DNS, CI/CD, SSL certificates — even the simplest static host assumes a baseline of technical knowledge. For a non-technical founder, marketer, or consultant, this is a brick wall.
• It's a project, not a workflow. Sharing one artifact means creating a repo, configuring a build step, setting up a domain. The overhead is wildly disproportionate to "I want someone to see this page."
• Updates are manual. Change the artifact in Claude, copy the new code, commit, push, wait for deploy. That's a software release process for what should be a single command.
• Access control is DIY. Want password protection? Build it. Want email-gated access? Build it. Want view analytics? Build it. Each feature is an engineering project.

4. Browser extensions (Backrun, etc.)

How it works

Browser extensions like Backrun add a "share" button to Claude's artifact panel. Click it, and the extension uploads the artifact to a third-party host, returning a shareable URL. Some extensions also add features like custom slugs or basic analytics.

What's good

• One click from the conversation. No copy-pasting code, no leaving the browser. The extension bridges the gap between "AI made this" and "here's a link."
• Low friction. Install once, click to share. The learning curve is near zero.
• Works today. Extensions can ship features without waiting for Claude or ChatGPT to add native support.

What breaks

• Third-party dependency. Your content lives on the extension provider's infrastructure. Their uptime is your uptime; their pricing changes are your problem; their shutdown is your dead links.
• Browser-only. Extensions work in the web browser. They don't work in Claude Code, Cowork, ChatGPT's desktop app, or any terminal-based AI workflow. If you work across multiple surfaces, you need a different solution for each one.
• Limited content access. Extensions can only see what the browser shows them. They can scrape the current artifact, but they don't have authenticated access to the AI's internal state — they're screen-scraping, essentially, with the fragility that implies.
• Security surface. A browser extension that reads and uploads your AI output has broad page access. You're trusting the extension author with everything your AI produces. Most are trustworthy; the trust model is still worth noting.

5. MCP connectors (Bauta, etc.)

How it works

MCP (Model Context Protocol) is the open standard that lets AI tools use external services. An MCP connector like Bauta gives your AI a "publish" capability — say "share this" in the conversation, and the AI publishes the artifact to a permanent URL with sharing controls. The workflow is conversational: no browser extensions, no copy-pasting, no leaving the chat.

What's good

• Works from the conversation. Publishing is a natural language command. The AI handles the upload, returns the link, and can modify sharing settings — all without you leaving the chat. "Make it public." "Add a password." "Update it." The AI understands these instructions natively.
• Permanent URLs with versioning. The link doesn't expire. Update the artifact next week and publish again — same URL, new content. Every previous version stays in history. When someone asks "what did this look like last Tuesday?" there's an answer.
• Sharing controls built in. Private, unlisted, email-gated, public — you choose per artifact, and you can change it any time. No DIY access control, no hoping the host offers what you need.
• Cross-client. MCP connectors work in Claude, Claude Code, Cowork, ChatGPT, Codex, and any MCP-compatible client. Set up once, publish from anywhere. Unlike browser extensions, you're not locked to one interface.
• Sandbox isolation. With Bauta specifically, shared content runs in an isolated sandbox domain. Content is scanned against Safe Browsing. Hosting is in the EU with a published DPA. There's a real security architecture, not just a terms page.

What breaks

• Requires a one-time setup. You need to add the MCP connector to your AI client. With Bauta, this takes about two minutes — but it's not zero friction. For a single throwaway share, this setup cost isn't justified.
• Account required. MCP connectors authenticate you so they can manage ownership and permissions. That means creating an account. If you want zero-account, anonymous sharing, this isn't the right tool.
• Newer paradigm. MCP is still gaining adoption. Not every AI client supports it yet (though the major ones do). If you work exclusively in a client that doesn't support MCP, this option isn't available to you.

The recommendation matrix

Different situations call for different tools. Here's a decision framework:

The trajectory: sharing is becoming publishing

A year ago, the question was "how do I get this out of the chat?" Today, the question is "how do I publish this properly?" That shift matters. When AI output was experimental — toy demos, proof-of-concepts — screenshots and quick shares were fine. Now that AI tools produce real deliverables, the sharing infrastructure needs to match.

The five methods on this list represent a spectrum from "I just want someone to see this for a second" to "I need this to be a permanent, controlled, professional publication." Most people need at least two of them — a quick method for throwaway sharing and a proper method for anything that matters.

Our recommendation: use Claude's built-in publish for the quick stuff, and set up an MCP connector for everything else. The two minutes of setup pay for themselves the first time you don't have to re-send a broken link, apologize for a stale version, or explain why your client proposal lives on someone else's domain.