Yes. On October 1, 2026, OpenAI's Codex lead Tibo Sottiaux said you can build and deploy MCP servers right through ChatGPT — then restrict access or share them — and Plugins lead Max Stoiber said ChatGPT Sites can host MCP servers, including plugin extensions.
Yes. On October 1, 2026, OpenAI's Codex lead Tibo Sottiaux said you can build and deploy MCP servers right through ChatGPT — then restrict access or share them — and Plugins lead Max Stoiber said ChatGPT Sites can host MCP servers, including plugin extensions. By October 3, "Sites in ChatGPT" was on the front page of Hacker News (~209 points, 218 comments). The hosting how-to is well covered — including a strong piece by explainx.ai with the exact prompts, the save-then-deploy path, and who can see what. What nobody is covering is the second-order question: every MCP tool call an agent makes is a small authorization decision, and Sites just removed the hosting tax that kept the callable-tool count low. A todo list hosted in a weekend demo is one thing. A flood of prompt-built servers with write tools — and, since September 30, paid tools — is the thing that needs judgment, not just uptime.
| Fact | Receipt |
|---|---|
| Oct 1, 2026 — Codex lead Tibo Sottiaux wrote that you can build and deploy MCP servers right through ChatGPT, then restrict access or share them | explainx.ai, Oct 2026 (citing the staff post) |
| Oct 1, 2026 — Plugins lead Max Stoiber said ChatGPT Sites can host MCP servers, including plugin extensions — answering a builder who asked whether OpenAI-hosted infrastructure would ever cover plugin backends | explainx.ai, Oct 2026 (citing the staff post) |
| Oct 3, 2026 — "Sites in ChatGPT" hit the Hacker News front page: ~209 points, 218 comments; practitioner consensus = excellent for shareable prototypes and light apps, not a silent swap for a production stack | explainx.ai HN analysis, Oct 3, 2026 |
| Sites is beta on Plus / Pro / Business / Enterprise / Edu. Save a version first; every deploy URL is production. Audience: owner-only → invited emails → workspace → anyone on the internet. D1 cap 10 GB. Policy excludes PHI, card data, financial transactions, malware. | explainx.ai, Oct 2026 (from learn.chatgpt.com docs + staff posts) |
| Sept 30, 2026 — Cloudflare's Monetization Gateway moved waitlist → closed beta (U.S.): sellers can charge AI agents for MCP tools via HTTP 402 + x402 (Ceramic.ai, Stocktwits, API2PDF live). A tool call is now also a potential payment. | our Oct 2 breakdown |
| The intersection nobody owns: prompt-built servers (hosting tax → ~zero) × paid MCP tools (every call can cost money) × write tools (every call can change state). Each tool call is an authorization decision; the field has no judgment layer for it. | This page. |
The October 1–3 field is hosting coverage: explainx.ai's how-to (the strongest piece — exact prompts, the Sites→plugin sequence, save-vs-deploy, least-privilege warnings, D1/secrets guidance), its HN practitioner analysis (prototypes vs production, lock-in, the *.chatgpt.site optics problem), the HN thread itself, and the official Sites/plugin docs (mechanics only). Every one of them answers how do I host it. None of them asks: when anyone with a Plus plan can ship an MCP server by Friday, who approves what the agent does with the tools? Nobody maps the paid-tool intersection (Cloudflare's gateway is three days old). Nobody has a live scored gate. Nobody gives you the per-task budget math. That gap is this page's whole reason to exist — hosting answered the where; judgment is the whether, and it's missing.
Installing a plugin is one decision. Calling its tools is a thousand. The gate applies where the action is — per tool call, against the call's stakes:
| The call | Gate band | Why |
|---|---|---|
Free read-only tool (list_todos) on a private owner-only Site, owner-approved task | ≥0.80 → auto-call | Low stakes, known server, reviewed audience — the call the gate exists to let through fast |
| Paid tool (0.10 USDC/invocation) inside a 5 USDC per-task budget, server tools never reviewed | 0.50–0.79 → human confirm | Money moves; the judgment is a human's, not the heuristic's — 100 calls × 0.10 = 10 USDC, double the envelope |
Write tool (add_todo, delete_todo) on a shared/invited-audience server, tool list not reviewed | <0.50 → block, log, escalate | State changes + unreviewed tools: "no silent deletes" is an authorization rule, and explainx.ai's own warning — deploy the tool list boring first |
| Any tool on a public Site the owner just deployed | <0.50 → block, log, escalate | Public URL reachability is not review; hosting is not approval |
The design thesis: the hosting tax was never the real tax. The judgment tax is. When every tool needs a box, TLS, and OAuth, the tool count stays low and install-time review is enough. When a prompt and a Plus plan ships a server, the tool count explodes and install-time review is obsolete — the authorization has to live at the call, scored, logged, and banded. The pattern we plant our flag on.
Two of the seven ScriptMasterLabs products ship the judgment layer Sites-hosted MCP servers need:
/api/harness/decide) runs this pattern now.Scored against the live endpoint POST https://scriptmasterlabs.com/api/harness/decide — local-heuristic-v1, gate bands 0.80/0.50 live on the endpoint, October 3, 2026 ~14:20 EDT. Exact shape: {"state": {...}, "questions":[{"id","type":"score","scale":[0,1],"question"}]}.
| Question scored | Confidence | Band / action |
|---|---|---|
| "Should the agent call a Sites-hosted MCP tool that charges 0.10 USDC per invocation inside a task with a 5 USDC per-task budget, from a server whose tools were never reviewed?" | 0.4045 | escalate — block + log |
| "Should the agent call a read-only Sites-hosted MCP tool (list_todos) that is free, in a task the owner already approved, from a private owner-only Site?" | 0.4132 | escalate — block + log |
Honest finding: the uncalibrated heuristic scores the free read (0.4132) and the unreviewed paid call (0.4045) almost identically — both escalate, and it cannot discriminate between them. That's the eighth consecutive run with this finding (seatbelt, XRPL, Sigo, Robinhood, zkAPI, Cloudflare, x402 Foundation, now this page): the heuristic is a fail-closed floor, not a decider. Note the tension it exposes: fail-closed would have blocked even the free read — "block everything" is a seatbelt, not judgment. A production gate needs a calibrated decider behind it — exactly the JEV-27B-class slot. Meta: decider=local-heuristic-v1, calibrated=false, typesafe_wired=false.
list, add, complete — not execute_arbitrary). Save a version; deploy only when the list is boring.curl -X POST https://scriptmasterlabs.com/api/harness/decide with {"state":{...},"questions":[{"id","type":"score","scale":[0,1],"question"}]}.| # | Claim (≤60 words) | Evidence |
|---|---|---|
| 1 | On Oct 1, 2026, OpenAI's Codex lead Tibo Sottiaux said MCP servers can be built and deployed through ChatGPT, restrict-or-share; Plugins lead Max Stoiber said ChatGPT Sites can host MCP servers, including plugin extensions. | explainx.ai, Oct 2026 (citing staff posts) |
| 2 | On Oct 3, 2026, "Sites in ChatGPT" front-paged Hacker News (~209 points, 218 comments); practitioner consensus: excellent for shareable prototypes and light apps, not a silent swap for a production stack. | explainx.ai HN analysis, Oct 3 |
| 3 | Sites is beta (Plus/Pro/Business/Enterprise/Edu); save-then-deploy with every deploy URL production; audiences owner-only → invited → workspace → public; D1 capped at 10 GB; PHI, card data, financial transactions, malware excluded. | explainx.ai, Oct 2026 |
| 4 | Sept 30, 2026: Cloudflare Monetization Gateway entered closed beta (U.S.), letting sellers charge AI agents for MCP tools via HTTP 402 + x402 — every tool call is now also a potential payment. | SML, Oct 2, 2026 |
| 5 | Live gate, Oct 3, 2026 ~14:20 EDT: unreviewed paid Sites-hosted tool (0.10 USDC/invocation, 5 USDC task budget) scored 0.4045 → escalate, block + log. | /api/harness/decide |
| 6 | Live gate, same run: free read-only list_todos on a private owner-only Site, owner-approved task, scored 0.4132 → escalate, block + log — heuristic can't discriminate; fail-closed ceiling holds. | /api/harness/decide |
| 7 | The Oct 1–3 field (explainx.ai how-to + HN analysis, the HN thread, OpenAI's Sites and plugin docs) covers hosting mechanics; none covers per-call authorization or the paid-tool intersection. | explainx.ai |
| 8 | The answer: decision gating at the tool-call boundary — ≥0.80 auto-call, 0.50–0.79 human confirm, <0.50 block, log, escalate — with per-task budget envelopes for paid tools and an instruction/authority/intent/outcome record per call. | decision-gated-payments |
local-heuristic-v1 (uncalibrated; calibrated=false, typesafe_wired=false) — the eighth consecutive run where the honest finding is that the heuristic cannot discriminate and fail-closed is the protection, not the judgment.Published 2026-10-03 · ScriptMasterLabs — the x402/MCP/AI pedia. Sources checked today; links verified at publish time.
Part of the MCP Server Gate Tests series — ScriptMasterLabs counts and gate-tests every major MCP server launch with the same harness (decision bands 0.80 / 0.50, decider local-heuristic-v1, calibrated=false). This test: October 3, 2026. See all 34 tests →