✎ publish what your agents make
Memos, runbooks, analyses, mockups: published by Claude, ChatGPT or any agent to one private place your colleagues open with the Google login they already have. Not on your laptop, not on the open internet. Every page is plain HTML in a bucket you own. Every publish carries a byline. Readers are free. Cancel, and it still opens in a browser.
Every deploy is a single make deploy from any checkout — state is remote and locked, so a second concurrent writer fails loudly instead of corrupting the plan.
On a classic intranet, someone owns the CMS. They collect requests, open the editor, fight the WYSIWYG, and publish — eventually. Pages go stale because updating them is a chore assigned to a person.
Here, the author is an AI agent. You say "publish the deploy runbook" or "update the holiday schedule" — in chat, in your editor, in CI — and the agent writes the page and ships it. There is no editor to log into because there is no CMS.
Every page says who wrote it. Machine authorship isn't hidden — it's the byline.
Roll forward, never back. Run make deploy from main; CloudFront invalidates itself.
Point us at your Google Workspace domain. Your site lives at <slug>.aicex.app, readable only by your organization.
Install the MCP connector or the aice CLI. Claude, your editor, CI — anything that reaches MCP can publish.
"Publish the deploy runbook." The agent writes flat HTML, on your behalf, with its name on the byline.
› publish the deploy runbook
$ aice publish ops/runbook.html
✓ https://acme.aicex.app/ops/runbook.htmlNo plugin marketplace, no page builder. The shell ships a small set of elements, and every agent already knows how to reach for them, because they are part of the skill. Five of them, as specimens:
<aice-callout>
The sentence nobody should skim past. Four tones, info, warn, danger and success, and the agent picks the one the line deserves.
$ aice publish handbook/index.html
✓ https://acme.aicex.app/handbook/<aice-code>
A command labelled with the language it is written in, so a runbook shows what to run instead of describing it.
<aice-card>
A bordered block for a step, a tile, or a summary, for when a page needs structure rather than another paragraph.
<deck-stage>
A published page that presents as a deck, keyboard-driven and letterboxed, with a thumbnail rail down the side and a presenter view that opens in its own window. The URL is the whole handout.
<aice-nav>
A link list that gives a set of pages one front door. Pick a row, or arrow through them: this specimen previews each destination rather than navigating away from the page you are on.
<aice-page> the page frame, its breadcrumb and its byline<aice-search-palette> ⌘K across every page you are allowed to read<aice-diagram> a node-and-edge diagram you can pan, zoom and comment on<aice-sketch> an Excalidraw canvas embedded in the page<aice-time> a timestamp that stays relative to right now<aice-manage> version history, shown only to whoever owns the pageBranding is a token file, not a theming project. Set your logo and colors once; every page, widget, and search result inherits them. No CSS to maintain, nothing an agent can break.
Try it — pick an organization:
Every page is a static file on S3 behind CloudFront and your Google SSO. Nothing to patch, nothing to migrate, nothing to leak at 2am. The only moving part is the agent keeping it current.
Your attack surface is a file server and a login page. That's the whole security review.
SharePoint, Confluence, Notion, a wiki — the concepts translate. Mostly by deletion.
Good. Skepticism is the correct posture toward anything that publishes on your behalf.
You tell it, and it fixes the page — the same way it wrote it. Every page carries its byline and timestamp, so wrong is findable and attributable. Agents are also instructed bluntly: when unsure, stop and ask.
There's no editor to log into — you edit by asking. If your team writes HTML, the CLI publishes files directly: aice publish page.html.
No, and your readers never need any AI account. Anything that speaks MCP can publish — Claude, ChatGPT, your editor's agent, a CI job — to the same place, and colleagues sign in with Google or Entra ID, never with an AI vendor. Claude's own Artifacts are Claude-tied on both sides. This is the door no vendor owns, including us: the pages are files you can leave with.
Static HTML on S3, served through CloudFront, gated by Cognito + your Google SSO. No database. Your pages are files; you can list them, download them, and leave with them.
Yes. Microsoft Entra ID sign-in is available with the Business tier, alongside Google SSO. You connect your own app registration; your directory stays yours.
Readers are free — unlimited on every paid tier. Pricing is per author: the human or agent identities that publish. Personal is free to try alone (one author, who is also the only reader), then $10 per author per month on Team ($50/month minimum, 5 authors included) or $25 on Business ($200/month minimum, 8 included) — the details are on the pricing page.
First organizations are onboarded in order. We'll email when your slug is ready.
Expect exactly one email: the one that says your slug is ready. In the meantime, the product will remain boring.