An alternative to sharing artifacts inside your organization

What Claude Teams artifact sharing is, and what this is

Artifact sharing inside a workspace is bounded by the workspace: viewers need to be members, and the page lives inside the product rather than at an address you control.

Publishing the artifact produces an ordinary URL. Whoever you send it to opens it in any browser with no account, no seat and nothing installed — and it is still unguessable, still gated if you want, and still on your domain if you want.

Where the two overlap

• Both keep the artifact away from the public internet.
• Both are one step from the conversation that produced it.

Where the defaults diverge

• Viewers need nothing. No account, no membership, no invitation flow.
• The page is at your address. Point a custom domain at it and the deliverable stops looking like somebody else's product.
• The link is stable. Re-run the agent, call update_site, and the reviewer refreshes the same URL.
• Access is per-page. Internal readers can get an email-domain gate; a client gets a password; a one-off gets burn-after-read.
• You can see it was read. Opens, unique viewers, how far people scrolled — and their comments come back anchored to the element they were about.

Side by side

Publishing the same page here

No account, no install — the anonymous route gives you a private link to compare against whatever you are using now:

curl -X POST https://read.botook.ai/v1/publish -F file=@index.html

With a key, from an agent, it is one tool call:

publish_html({
  content: "<h1>Q3 review</h1>…",
  expires_in_hours: 168,
  password: "optional"
})

The second revision is where it shows

Agent loops do not produce one page, they produce fourteen versions of one page. update_site writes the new bytes behind the existing token, so the address you already sent out keeps resolving to the current draft:

curl -X PUT https://read.botook.ai/sites/SITE_ID \
  -H "Authorization: Bearer otr_live_YOUR_KEY" \
  -F file=@report.html

Three gates, stacked or separate

Plus expiry — set_expiry in hours, or never — and burn-after-read, which closes the page permanently after the first successful open.

What you can hand it

The upload primitive is the same on every route, so nothing has to be converted before it is sent. A single document, a folder, or a drawing all go through one call:

Markdown is rendered when it is published rather than when it is read, so the reader, the preview card and /raw are all looking at the same stored document. Notebooks, diagrams and slide decks travel as whatever they export to — this hosts the export, it does not run a converter over your source.

What the reader sees

No account and no install: a reader opens the link and the page is simply there. Everything optional is off unless you asked for it.

• Nothing to sign into. The token in the URL is the credential — until you stack a password or an email-domain gate on top of it.
• An optional toolbar. With set_agentation on, a reader can select any element and leave a note anchored to it, which you read back with list_feedback and close with resolve_feedback.
• Optional reactions. A separate switch from feedback, because a reaction is an opinion and a note is a task.
• A link preview. Pasting the URL into a chat app renders a card from a real screenshot of the current revision, unless a gate is up — a page behind a password is never photographed.

What you learn afterwards

Sending a link and hearing nothing back is the normal failure of every tool in this category. Read-side numbers are collected per page and per revision without cookies or a third-party tag:

• Views and unique readers, and when the page was last opened.
• Dwell time and scroll depth, plus whether anyone reached the end.
• A ten-segment heat map showing where attention stopped.
• Unresolved notes and reaction counts, carried on the site object itself.

curl https://read.botook.ai/sites/SITE_ID/engagement \
  -H "Authorization: Bearer otr_live_YOUR_KEY"

Moving a page across from Claude Teams artifact sharing

Copy the artifact's HTML, publish it, and send the link. From then on, ask the agent to publish directly and skip the copy step.

1. Get the HTML — export it, view-source it, or ask the agent that wrote it for the file.
2. Publish it: paste it on the home page, drop it in the dashboard, or POST /sites with your key.
3. Set the access rule the page deserves: password, email-domain gate, or nothing but the token.
4. Send the new link, and let the old one expire rather than deleting it out from under anyone.
5. Wire the agent up once — see integrations — so the next revision never passes through your hands.

Getting your pages back out

The question worth asking before you move anything in: how hard is it to leave? Every published document can be read back in full, by the key that owns it, in the exact bytes that are served — plus a plain-text rendering for whatever needs to re-read it as prose:

curl -s https://read.botook.ai/sites -H "Authorization: Bearer otr_live_YOUR_KEY" | jq '.sites[].id'
curl -s https://read.botook.ai/sites/SITE_ID/content -H "Authorization: Bearer otr_live_YOUR_KEY" -o page.html
curl -s https://read.botook.ai/raw/UNLISTED_TOKEN -o page.txt

Nothing is stored in a proprietary container and nothing is proxied from somewhere else — a published page is the file you handed over, in object storage, addressed by a token. A loop that wrote its output here can walk its whole list_sites result and pull every page down in one pass, which is the only export guarantee that means anything.

Where the bytes actually live

This is a self-hosted deployment on Cloudflare: the request is handled at the edge, the document sits in object storage, and the metadata is one row in a SQL database. That has three consequences worth knowing before you compare it with a managed product.

The trade is explicit: you inherit the operational surface — the domain, the storage bill, the upgrade — in exchange for nobody else being able to change the terms under you. Self-hosting covers what that actually involves.

Evaluating both at once

You do not have to pick before you know. Publish the same document to both, send the pair to somebody, and compare the two things that only show up in real use: what the recipient has to do before they can read it, and what you can find out afterwards.

1. Publish it here anonymously — no account, no project to create, one curl.
2. Publish the same file to Claude Teams artifact sharing the way you normally would.
3. Send both links to the same reader and watch which one they open without asking a question.
4. Change the document, and update each one. Note whether the link you already sent still works.
5. Check what each side tells you afterwards about whether it was read.

Step 4 is where the difference usually lands, and it is the step people skip when comparing on a feature table. Publishing directly from the agent that wrote the page only pays off if the second revision costs nothing.

When Claude Teams artifact sharing is the better answer

For a conversation among colleagues who all have seats, internal sharing keeps the context attached to the artifact, which is worth something. Publish when the audience is outside.

This page would be worth less if it pretended otherwise. The comparison that matters is between defaults, and defaults are only right relative to a job.

Questions people ask

Does the reader need Claude?

No. They need a browser. That is the entire requirement.

Can I restrict it to my company anyway?

Yes — the email-domain gate does exactly that, and it works for people who have no seat in your AI vendor's product.

What does it cost?

Nothing — it is self-hosted on Cloudflare, so pricing is a description of your own bill rather than a set of tiers.

Which agents can publish to it?

Anything that speaks MCP over HTTP: Claude Code, Codex, Cursor, the Claude.ai connector — or plain curl from a script.

Where is the full reference?

The docs: every endpoint, all twelve tools, the upload formats and the defaults applied to each page.