The Stacktree blog.

Posts on the publish primitive, MCP servers, agent-loop hosting, and what changes for engineering teams when AI agents emit HTML at human-scale rates.

Get started free

No card · 3 pages free · about a minute

What does the Stacktree blog cover?

Long-form posts on AI agent infrastructure: the Model Context Protocol, the publish-primitive pattern, agent-loop semantics, and the operational implications of agents emitting HTML at the rate humans write commits. Every post is dated, attributed, and updated when the underlying landscape shifts — the same EEAT contract that applies to the rest of the site.

Latest posts

post The first machine to buy past our entry price One wallet paid $1 for an API key, published a page, then bought permanence and higher limits. It is the first of seven x402 identities to buy anything above the door price. The receipts are on Base, and here is what we can and cannot tell you about it.
post Page engagement analytics, without recording the reader We wanted to show how a shared page is read: time on page, scroll depth, where attention landed. We did it without session replay, keystrokes, or identity. The design, and why aggregate was the point.
post What is the Notion HTML block? Uses, limits, sharing Notion 3.6 adds an HTML block: agents build interactive HTML, calculators, quizzes, diagrams, right on the page. A plain reading of what's announced, what's not documented yet, and the sharing gap it leaves.
post Cloudflare Monetization Gateway: x402 at the edge, explained Cloudflare announced a Monetization Gateway: charge agents for pages, APIs, datasets, and MCP tools over x402, enforced at the edge. What it does, what it leaves out, and notes from running agent payments in production.
post Throwaway dashboards: where the agent-built one lives Agents now build a dashboard on demand and throw it away. The pattern works, but the HTML still needs a home. How to host an agent-built dashboard on a private link that expires on its own.
post Why AI agents need a publish primitive Agents write HTML at the rate humans write commits. The publish step deserves its own verb — not a workflow grafted onto a static-site host.
post What is an MCP server? A developer reference (2026) MCP is the protocol every AI assistant now speaks: Claude, ChatGPT, Cursor, Codex. How MCP servers work, what they replace, and how to build one in about 50 lines.
post Private-by-default HTML hosting is the new normal Agents emit a lot of HTML with real data in it. Public-by-default hosting is no longer the safe baseline. Here is what the new baseline looks like.
post Do AI agents leak secrets? Yes. The publish step is where it becomes public. Agents assemble pages from whatever is in context, and context routinely contains live credentials: the bearer token that fetched the data, the key in a pasted config. What a pre-flight secret scan catches (API key prefixes, JWTs, Luhn-checked card numbers, SSNs, emails), why the MCP publish path blocks by default, what a regex net honestly cannot promise, and what to do when a key already shipped.
post MCP 2026-07-28 spec: every breaking change, with fixes Every change in the final MCP 2026-07-28 spec and how to migrate: stateless core, MCP Apps, Tasks, elicitation, sampling, caching. What breaks, what stays compatible, and the fix for each.
post MCP Apps made HTML the official UI of agents. Where does that HTML live? MCP Apps (SEP-1865) is final: Claude, ChatGPT, VS Code, and Goose all render server-shipped HTML in a sandboxed iframe. The extension standardizes emitting interactive HTML and is deliberately silent on hosting it beyond the session. What it covers, its lifecycle, and where the publish step fits.
post MCP tools vs resources vs prompts: a decision table Tools act, resources read, prompts guide. A decision table for which MCP primitive to use, the mistakes that come from picking wrong, and what the 2026-07-28 spec changed.
post Skills vs MCP vs API: when you need each (2026) API is the surface, MCP is the interface, a skill is the expertise on top. The three layers of an agent-facing product, explained with Stacktree, which ships all three.
post What is MCP elicitation and sampling? Elicitation lets an MCP server pause and ask the user for input mid-call. Sampling lets it ask the client model to reason. Together they turn servers from passive endpoints into active participants.
post Codex Sites is now ChatGPT Sites: what the docs don't say Codex Sites went GA as ChatGPT Sites on 9 July 2026: every paid plan, public publishing, custom domains. Where the official docs stop: gates, expiry, API access, and private alternatives.
post Can you make a Codex Site public? Plan availability and access (June 2026) No public or link-only mode is documented for Codex Sites: all three access modes require an OpenAI workspace seat. Which plans have it, what OpenAI says is coming, and how to get a public link for agent output today.
post Can a Codex Site have a custom domain? No custom-domain support is documented for Codex Sites: a deployed Site renders only at an OpenAI-hosted, workspace-scoped URL. Why it is built that way, and how to put agent output on your own domain today.
post What are Claude Live Artifacts? Access, plans, and the sharing gap A plain reading of Anthropic docs: Claude Live Artifacts are persistent Cowork dashboards that re-query your apps and refresh on open. Paid plans, desktop only, personal and local at launch, not yet shareable. The private-link gap Stacktree fills.
post Artifacts in Claude Code: what they are, who can view, sharing limits (2026) A plain reading of Anthropic's launch: Claude Code turns a session into live pages that update in place. Org-only and cannot be made public; Team and Enterprise beta. The external-sharing gap Stacktree fills.
post Claude artifact unpublished and cannot be republished? Why, and the fix Unpublishing a Claude artifact is permanent for that version: the link dies, the Publish button deactivates, persistent-storage data is deleted, and republishing mints a new URL. Why it happens, and how to keep a link that stays put.
post Stacktree is an n8n node now n8n-nodes-stacktree passed n8n's review and is live: any workflow can publish HTML to a private link, update it in place, gate it, and read viewer feedback. Works as an AI Agent tool.
post Slack renders HTML attachments now. Rendering still is not sharing Slack is starting to render attached HTML files inline instead of a download or raw code. Useful, and rolling out. But a rendered file in a channel has no link, no access control, and no life outside the workspace. How to turn it into a private link.
post The end of per-seat pricing: agents don't have seats (2026) Per-seat pricing assumes one human per seat. AI agents do not log in or take seats, so the model breaks. Original data on what static hosts charge per seat today, why the category is moving to flat and usage pricing, and where it leaves teams running agents.
post x402 and MPP, explained by building a service that takes both What x402 and MPP are, how an AI agent pays with each, and how we built a multi-rail service with agentcash where the payment turns into a site the agent or a human can then claim and manage.
post MPP in production: how AI agents pay Stacktree over Tempo An early, live example of MPP, the agent-payment standard from Stripe and Tempo. AI agents pay Stacktree over MPP or x402 on one endpoint, with no account and no card. How it works, why the onchain method is permissionless, and the on-chain receipt.
post An AI agent paid us $1 to provision itself: x402, end to end A working x402 example on Base mainnet: an agent hits a 402, signs an EIP-3009 USDC authorization, pays $1 with no human and no card, gets a persistent API key, and publishes a live page. The whole loop, the real transaction, and the bug that cost two attempts.
post MPP vs x402: you do not have to choose MPP and x402 are usually framed as crypto versus Stripe. That is a false binary: MPP’s evm method settles the same EIP-3009 USDC on Base that x402 does. What actually differs, what is identical underneath, and why we accept both on one endpoint, with a real on-chain MPP payment as proof.
post What is Shopify's Quick? What it proves about AI-era hosting Shopify built an internal host for AI-made HTML: drop a folder in, get a company-only URL. Half the company used it. Why constraints won, and the public version.
post What is Vercel Drop? Limits, privacy, and when to use it Vercel Drop (June 2026) deploys a dragged file or folder with no Git or CLI. What it does, its documented limits, and when an artifact wants a different host.
post ChatGPT supports WebMCP now OpenAI added WebMCP to the ChatGPT desktop app browser, Work, and Codex, with no origin trial token. What document.modelContext is, how to register a tool, and the annotation detail that decides whether ChatGPT asks permission first.
post How to export a coding agent session — and make it worth reading YC now asks applicants to attach a coding agent session. Exporting is easy (/export in Claude Code, transcripts in Codex, a button in Cursor); the export is a 4,000-line wall. How to get it out of each tool and turn it into a page someone reads.
post is-agentic explained: what the agent-readiness score measures, and the fix that took us from 93 to 100 We ran Vercel's new is-agentic scanner against stacktr.ee: 93 at baseline, 100/100 a day later. The gap that mattered most was invisible to humans: 99KB of inlined CSS pushing the first heading past byte 100,000. Every fix, ranked by what it was actually worth.
post skills.sh indexing explained: why your skill is not listed, and what actually breaks The agent-skills leaderboard has 191 open "request indexing" issues. We shipped a listed skill, hit two silent failures on the way, and wrote down what the index actually does: parsing, repo layout, ranking, and the diagnostic checklist we wish existed.
post Grok Bot explained: what it is, what it costs, and what it can reach xAI's Grok Bot is an AI teammate with its own cloud computer that signs into your apps by driving a browser. What it does, what it costs, why it needs no integrations, and what we found when we asked one to publish a page. Tested hands-on.
post What is ChatGPT Sites? Plans, cost, and the 3 real limits What ChatGPT Sites is, who can create one, what it costs, and the three limits we hit testing it from London: no gate a client passes without an account, no API, and beta usage caps.
post How to share a ChatGPT Site with a client ChatGPT Sites now has three audiences: your workspace, named people you invite by email, or the whole internet. The invited client still signs in with a ChatGPT account. What the options are after the September 2026 private-sharing update, and when you need a different tool.
post ChatGPT Sites private sharing: who can open an invited Site ChatGPT Sites can now be shared privately with named people outside your workspace (11 September 2026). What the invite does, why the viewer still has to sign in with a ChatGPT account, and the account-free option when the reader is a client.
post ChatGPT Sites usage limits: what the 50% warning means ChatGPT Sites is a public beta with plan-specific limits on sites, storage and visits, and no way yet to buy more. What happens at the limit, what users report, and how to keep finished pages online without spending Sites quota.
post ChatGPT Sites examples: 40 real sites from the first three months Forty ChatGPT Sites people actually built, from OpenAI's showcase and the community awesome-list thread: games, learning tools, creative experiments, personal sites on custom domains. What the pattern says, and where a client deliverable belongs.
post How to delete a ChatGPT Site, and whether it frees up usage Deleting a ChatGPT Site takes four clicks and cannot be undone. Whether it frees your beta usage is less clear: one user deleted twelve Sites and stayed blocked. The steps, what deletion does and does not free, why usage does not reset weekly, and the levers that reduce it.
post Codex keeps publishing to ChatGPT Sites. Here is how to stop it Codex reaches for ChatGPT Sites whenever a task produces something viewable. The AGENTS.md line that stops it, why the word "website" triggers it, and how to give Codex an explicit publish tool that is private by default.
post Do small agencies actually use a client portal? What agency owners say Mostly no. Agency owners and freelancers compared notes this week and the answers were email, Drive, Slack and WhatsApp, plus one rule: the portals that survive are read-only views of work the team already tracks. What clients actually want, and the lighter pattern.
post Claude artifact too large to share: the error, the limit, the fix Private viewing works, public sharing fails with "This version can't be shared publicly." Anthropic documents no size limit, but large artifacts hit one around 16 MiB. What causes it, how to shrink the page, and how to share the full-size version on a private link.
post Monthly client reports without the deck: the page that replaces the PowerPoint An agency spending ten hours a month on report decks asked for a better way. The answers: drop the static deck, give each client a live view and a short written narrative, and separate the numbers pull from the human "why". How to do that as one private page that updates every month.
post Your client didn't skip the proposal. They never opened it. Thirty days of first-party reading data from 375 client pages: once a page is opened, 84% of sessions get halfway and 70% reach the bottom. The bigger loss is earlier: 39% of recent pages were never opened at all. How we measured it, the honest limits, and four ways to get a proposal or report opened.
post How to send a proposal that opens on its video Give a proposal or report a one-minute page video, then send it so the video is in the email itself: a poster picture that shows the headline, and a link that opens straight into the video. Step by step in the Stacktree dashboard, how to see who watched, and when to send a Loom as well.
post Google Vids turns docs into videos. When to send something else. Google Vids can now turn a Google Doc, PDF or Word file into a narrated video summary. What it does, who has it, where it fits well, and why a client deliverable you send outside the company may need a different kind of video.
post Feedback is not approval: how agencies close the client feedback loop Client feedback scatters across email, design comments and chat, and agencies lose clients over it. The fix is one place tied to the exact words and version, a clear close on each comment, and a separate step for sign-off.
post awesomesites.ai: the ChatGPT Sites gallery, what is in it, and how to get listed Kath Korevec of OpenAI launched awesomesites.ai on 13 September 2026: 80 ChatGPT Sites with previews and upvotes, nominations by posting with #awesomesites on X. We went through all eighty. What people build (games, 3D scenes, music, maps, learning tools), the two on custom domains, what is absent, and how to get a site in.
post Claude artifact "Page not found": why the link a client sent you is dead, and what to ask for A Claude artifact link that opens as "Page not found" was almost always shared to the sender's organisation, not to you. The three causes behind one screen, why they hit contractors and agencies hardest, the message to send back, and why "Anyone with the link" is the wrong fix.
post Every Stacktree page can now answer questions about itself Ask this page is live: an opt-in question box on any published page. The answer comes from the page only, streams in the page's own type, and lights up the passage it came from. Where the idea came from, the three decisions that stop it being a chatbot, and the Markdown twin every page now has for the reader's agent.
post Meta opened Muse to connectors. Here is what the form asks for, and what we filed. The Muse Connector Platform takes third-party connectors for Meta's personal agent: describe the product, pass a functional, security and legal review, appear in a directory. The three-step form field by field, what a connector is technically, and what Stacktree submitted on 19 September 2026.
post Claude Code wants to publish an artifact: what "new artifacts stay private until you share them" means The permission prompt says the tool uploads the file to claude.ai, that new artifacts stay private until you share them, and that a re-publish keeps its sharing settings. What each clause means, when Claude Code publishes without asking, who can see the page on each plan, and how to get it to someone outside your organisation without making it public.
post How to share a Claude Doc or Claude Slides deck outside your organisation Claude Docs cannot be shared outside a Team or Enterprise organisation, and every shared doc needs a Claude account to open. The export routes that get a doc or a deck to a client, contractor or board as a private link with no account, and the one-sentence version from Claude Code.
post What is /design in Claude Code? Tested at launch The /design research preview shipped in Claude Code 2.1.234 with no docs. We ran it: artboards land as .dc.html files in your repo and the canvas publishes as a claude.ai artifact.
post What is Cloudflare Drop? 60 minutes to claim, then it is gone Drag a folder or zip, get a live workers.dev site with no account. You have 60 minutes to claim it or Cloudflare deletes it. The file limits, the public-by-default URL, and the terms you accept.
post Claude artifacts now share publicly and edit multiplayer (2026) Claude Code shipped public sharing and multiplayer editing for artifacts on 13 July 2026. What it does, the three things it still cannot do (private gating, your own domain, non-Claude agents), and where a publish primitive fits.
post How to add x402 payments to your API: what it actually took The build guide behind our live x402 endpoint: a real 402 (v1+v2), facilitator settlement, idempotent minting, SKU pricing that cannot bleed you, discovery listings, and the mainnet gotchas.
post x402 in production: 30 days of a live paid endpoint Real numbers from a live x402 endpoint: settled payments on four rails, x402scan vs the application ledger, what buyers did after paying, and the gotchas that cost live transactions.
post What happens when an AI agent loop hits a paywall? Loop engineering says design loops that ship while you sleep. A 402 at 3am is a stop condition you did not design. Two live rails let a loop pay and keep going.
post What AI agents actually publish: data from our first 80 sites Early aggregate data from Stacktree: 99% of agent-published pages are a single self-contained file, 86% are under 100KB, and over half are password-gated. Agents emit private artifacts, not projects.
post WebMCP in production: how the Stacktree dashboard speaks it WebMCP lets a page register tools that in-browser AI agents can call. Stacktree's dashboard now exposes its site actions as WebMCP tools, behind one cmd-K palette.
post Our homepage publishes for you now: an AI agent builds and ships a page, no login We put a WebMCP publish tool on the public site. A browsing AI agent, or you in plain English, can generate an HTML page and ship it live from stacktr.ee with no sign-in.
post Clients never log into your portal. The login is the reason. The recurring complaint about client portals is not a missing feature, it is that the client never went in. The category knows, and its own published fix, put the invoice behind the login, gives the game away.
post One folder per client is right. The client cannot open a repo. Agencies are moving client work into repos, one folder per client, with the agent skills sitting beside it. The structure is correct. It runs out at the handover, because a repo is a workshop and the client has no key.

References

Sources and further reading

Model Context Protocol — spec ↗ The standard that underlies most of what we write about.
Karpathy — agents output HTML ↗ The thread that crystallised the agent-output-as-HTML thesis.

Have a topic we should cover?

Reply to any post or open an issue. The blog is small and curated, not a content factory.