Market study — demand first, then product

Date: 18 September 2026
Role: Project manager, CryptoBook pivot
Constraint: build is PAUSED. This file is a go/no-go on demand, not a survival spreadsheet. Do not write application code. Do not touch nginx.
Failed prior process: DECISION.md invented Slip by stitching leftover patches after Hallpass, Permitfile, Workslip, Rehearse, SchemaPin, and every quantum SKU died a legal/security/finance/competition/adoption filter. That is compliance-shaped product design. It is not a market study. Treat it as a cautionary tale.

1. Method: demand first, then product

TAM slides are not demand. Bain’s “$100B US SaaS opportunity from agentic glue work” and MarketsandMarkets’ “$19.33B agentic AI in 2026” describe analyst envelopes. Nobody with a Visa is buying a percentage of those.

A job is real in this study only if all five are true this week:

1. A named person already feels it. Not a CISO slide. Not “governance should exist.”
2. They already pay someone, or they have a documented invoice/horror story and a shipping product with a public price they could swipe into.
3. The incumbent is named. “Nobody does this” is usually “nobody pays for this.”
4. The object they buy is the work, not a protocol, a YAML, a signed hash, or a TAM cell.
5. 100–250 paying accounts at $19–99/mo is a plausible 12-month outcome from cryptobook.space (a leftover crypto-social domain) with content + playground, no sales team.

What this study will not count as evidence:

• Survival-rule patches (“legal SURVIVE, security SURVIVE, one FIX”).
• “A + B + C and nobody ships all three” aggregators. That is how Hallpass became Slip.
• GitHub star counts, CVE counts, OWASP Top-10 lists, MCP SDK download numbers. Those are protocol heat, not card heat.
• Product Hunt upvotes without a pricing page that people actually reach.

Primary sources opened 18 Sep 2026 are listed at the end. Numbers from pages not opened are not used.

2. Where money is actually moving in September 2026

2.1 Consumer and prosumer AI (card, this week)

Menlo Ventures, 2026 State of Consumer AI, GlobeNewswire 16 Sep 2026 (survey of 5,067 US adults, July 2026):

• Global consumer AI spend $40B in 2026, from $12B in 2025. User base 1.8B → 2.0B (+11%). US adult usage 61% → 64%.
• 55% of US AI users pay for at least one AI product. 46% of payers spend more than last year.
• 14% of payers spend $100+/month and account for ~60% of US consumer AI spending.
• Daily use: 19% → 25% of US adults.
• Agents: 41% of US AI users have tried an AI agent; 24% use one regularly; 32% have let AI take an action at least once without final approval. Regular agent users: 92% pay for AI vs 55% of AI users overall.
• Trust now outranks convenience: accuracy 45%, trustworthiness 40%, security/privacy 36%, ease of use 32%.

That $40B is ChatGPT / Claude / Gemini / Cursor / Copilot / image-video, not MCP gateways. Growth is intensity among existing payers, not a new user wave.

What those payers swipe, verified on vendor pages this date:

Cursor’s own pricing page is the cleanest “person with a card this week” object in the whole agentic stack: Hobby free → Pro $20 → Teams $40/user. Claude Code is not a separate SKU; it is a feature of Pro/Max. That is where the coding-agent money sits.

2.2 What builders do every day (not what they pay a third party for)

LangChain State of Agent Engineering, 12 June 2026 (1,340 responses, Nov–Dec 2025; 63% technology, 49% of orgs <100 people):

• 57.3% have agents in production; another 30.4% developing with a deploy plan.
• Top barrier: quality 32%. Latency 20%. Cost dropped vs prior year.
• At 2k+ employee orgs, quality still #1; security #2 at 24.9%.
• Observability: 89% some form; 62% step/tool tracing. In production: 94% / 71.5%.
• Offline evals 52.4%; online 37.3%.
• Daily write-ins: coding agents dominate (Claude Code, Cursor, GitHub Copilot, Amazon Q, Windsurf, Antigravity). Then research. Then custom LangChain/LangGraph.
• Forced primary use case: customer service 26.5%, research/data 24.4%, internal workflow 18%.

Read-through for a 1-VPS indie: production is real, coding agents are the habit, quality is the killer, observability is already table stakes (do not clone Langfuse), and cost is no longer the #1 production story. Personal bill shock still exists (see §3), but the vendor dashboards are absorbing it.

2.3 Enterprise agentic money (not our buyer)

This is where the governance rush lives. It is real. It is not a $19–99 card.

• Okta Agent SSO GA 24 Aug 2026. Bundled at no extra cost into core Okta SSO (~20,000 customers). Cross App Access agents register in Universal Directory; short-lived tokens replace static keys. Okta for AI Agents (GA May 2026) is a separate subscription. Source: Okta press release.
• IBM watsonx Orchestrate / AgentOps Agent — sales-led control plane (prior research: Essentials from ~$530/mo). Not opened again this pass; not a card SKU.
• Broadcom AgentMinder — VMware Private AI Cloud, no list price.
• Salesforce Agentforce — bundled into Core/Advanced/Max editions at hundreds of dollars per seat (press 3–4 Sep 2026). Procurement.
• Customer-support agents (Intercom Fin, Zendesk, Sierra, Decagon) — outcome or conversation pricing, often sitting on an existing helpdesk contract.

A 1-VPS product that looks like “governance” will be compared to free Okta Agent SSO and then ignored.

2.4 Product Hunt, this month — what the builder crowd upvotes

Best of Product Hunt, September 2026 (opened 18 Sep 2026):

Nothing in the September 2026 PH top 10 is an MCP gateway, a blast card, a signed receipt, an observability clone, or a quantum classroom.

PH is not invoices. It is still a useful negative: the category Slip wants to occupy is not even on the board. Adjacent caution: AI Toolbox 3.0 won #1 Product of the Day earlier this month with 410 upvotes; the maker wrote that 23 people reached the pricing page. Upvotes ≠ cards.

2.5 Quantum is dead for a 1-VPS SaaS (reconfirmed)

Opened this date:

• IBM Quantum access plans: Open free, PayGo $96/min, Flex $72/min (400 min floor), Premium $48/min (5,200 min floor). The product enterprises pay for is minutes on a real QPU. ibm.com/quantum/pricing
• IBM Classroom Accounts (28 May 2026): 5–100 students, no credit card, real QPUs, 10 min/student/month, 365-day validity, paired with free Learning modules. IBM Quantum blog
• Composer is a free browser circuit builder on the same platform.

This VPS is a ≤32-qubit numpy box. IBM already gives free real 100+ qubit hardware + unlimited simulators + a classroom SKU with no card. Zapata (software-only application layer) is the cautionary tale from prior research: $0 revenue after ~$118M. Do not ship quantum. Do not costume the agentic product as quantum.

3. Jobs a person with a credit card will pay $19–99/mo for this week

Filter: named incumbent, public price, evidence they swipe. Not “would be nice.”

Job 1 — “Write my code with an agent” (the actual card)

• Buyer: individual developer; 2–15 person product team.
• Pays today: Cursor Pro $20 / Teams $40/user (cursor.com/pricing); Claude Pro $20 / Max from $100, Claude Code included (claude.com/pricing); GitHub Copilot on a GitHub seat.
• Evidence: LangChain daily-agent write-ins are these names. Cursor’s page is self-serve card. Claude Pro is self-serve card. Menlo: regular agent users 92% pay.
• Wedge for us: none. We are not Anysphere or Anthropic. A “better Cursor” on cryptobook.space is a joke.

Job 2 — “Be my general assistant / research / writing” (the other actual card)

• Buyer: the Menlo 55% who already pay.
• Pays today: ChatGPT Plus $20 / Pro $100–200 (not re-opened this pass; treated as known consumer SKU), Claude Pro $20, Google AI Pro ~$19.99.
• Wedge for us: none. Model labs own it.

Job 3 — “Do my inbox / Slack / meetings as a teammate”

• Buyer: operator, founder, sales/CS individual.
• Pays today: Lindy Plus $29.99/user/mo, Pro $99.99, Max $199.99. Credits pause rather than surprise-bill. Slack join = 7-day seat trial. lindy.ai/pricing
• Incumbent: Lindy, plus ChatGPT / Claude connectors, Superhuman, Granola (meeting notes; named as an Okta XAA integration).
• Wedge for us: none. Needs thousands of integrations, computer-use, Slack-native UX. Not a 1-VPS weekend.

Job 4 — “See traces of the LLM app I already shipped”

• Buyer: team with a production LLM/agent application (not a Cursor user).
• Pays today: Langfuse Core $29/mo (100k units, unlimited users, 90-day). Canva, Twilio, Ramp, Khan Academy listed on the pricing page. MIT self-host is $0. langfuse.com/pricing. LangSmith Plus $39/seat. Helicone still live in maintenance mode after Mintlify (3 Mar 2026, helicone.ai/blog/joining-mintlify) — 16k orgs historical, not a hole.
• Wedge for us: forbidden. BRIEF and every challenger kill a Langfuse clone. 89% already have observability.

Job 5 — “Tell me which customer/PR ate the OpenAI + Cursor bill”

• Buyer: small team with a four-figure AI invoice they cannot map.
• Pays today: SuperPenguin Growth $30/mo (up to $5k managed spend), Pro $200. Free under $2k. SDK is not a proxy; Mac app for Cursor/Claude Code/Codex. superpenguin.ai
• Incumbent: SuperPenguin (YC); provider consoles; Vantage/CloudZero for cloud; OSS clones exist (publicly advertised against SuperPenguin’s $200 tier).
• Wedge for us: occupied. Also not 60-second without a Mac app + GitHub.

Job 6 — “Charge for the AI thing I built this weekend”

• Buyer: indie who shipped with Cursor/Claude and now needs auth + Stripe + tax.
• Pays today: they pay Stripe / a Merchant of Record, not a $29/mo “permit.” September PH #1 is tiun. (auth + billing + payments + customer DB). Creem 2.0 also launched as “seller of record” (3.9% + 40¢, no monthly fee) — seen in PH recap, not re-opened as a pricing page this pass.
• Wedge for us: legal kill. Holding customer funds or acting as MoR is money transmission. BRIEF hard constraint 3.

Jobs that look hot and do not clear a card

These were the previous pivot’s entire idea list. Demand evidence is absence.

The $1,263 Cursor bill is the strongest pain quote in this file. The founder’s response was to build an OSS MCP, not to subscribe to a hosted permit plane. HN did not care (6 points). Cursor’s own pricing page now documents on-demand usage and an Admin Dashboard. Claude settings and Anthropic console caps exist. The incumbent for “stop the runaway loop” is the vendor the user already pays.

4. What is crowded (do not enter)

Confirmed this date, not a slide:

If the homepage, SDK, or pricing looks like any row above, the pivot is already dead.

5. Honest verdict on Slip / Hallpass / blast-card

KILL. Do not ship. Demand evidence, not survival-rule evidence.

Hallpass (hosted MCP URL + English + $ cap + receipts)

Killed three times already by security (key honeypot), competition (Permit.io), and adoption (policy editor is homework; stdio never hits the URL). Demand adds a fourth kill: the category leader’s own launch did not move HN (2 points). Cursor and Claude already nag. OpenAI/Anthropic/Cursor already cap dollars. A third-party English textarea in front of a proxy is not a job anyone is paying for this week.

Blast Card (paste MCP → shareable scare)

Adoption invented this so something would pass the 60-second test. CodeFlowOps already ships the 30-second grade for free. ShieldMCP already ships the “paste your mcp.json” aha; the paid conversion is a $49 PDF, and Product Hunt treated it as a ghost (5 upvotes, 0 reviews). That is the demand ceiling for this object: a one-shot report, if that. There is no $19–99/mo habit. A public /c/ card is a tweet, not a company. SSL Labs taught this lesson a decade ago.

Slip (Hallpass with the poison cut out: compiler + local PEP + hashed receipt + blast card)

This is the Frankenstein DECISION.md blessed. Competition said it survives only if it refuses the gateway category. Finance needed six named stretch levers to clear $5k MRR. Security forbade holding keys.

Demand does not care about the architecture split.

• The receipt as a product: cMCP Show HN, 9 points, comment thread is “your launch comment was AI-generated.” Obsigna / ProofAgent / IETF draft-sahu exist as protocols. Nobody’s client asked for a cryptographic receipt this week. Freelancers send a git diff.
• The compiler: English → YAML is a weekend in front of Preloop/Permit YAML. Not a SKU.
• The local PEP: mcp-doorman, satgate-proxy, Cordon, Claude native policy, Cursor allowlists. npx is free.
• The paywall (unwatermarked receipt + 90-day deny inbox): a paywall in search of a buyer. Finance’s own base case was ~$4.0k MRR assuming the playground converted — and that assumption was never demand-tested.

Survival math is not willingness-to-pay. Slip exists because the previous PM was not allowed to return an empty winner. Operator override: that is invalid.

6. Alternative jobs, ranked, then killed

Score each on acute pain × WTP × uncrowded × 60s start × 1-VPS × US-legal. Kill on any fail. Rank is “least dead first,” not “ship.”

Alt 1 — Team coding-agent spend, cost-per-PR

KILL (crowded)

• Pain: real. l6e founder, $1,263 Cursor month. Cursor forum threads this month about usage hitting 100% mid-cycle.
• WTP: SuperPenguin already prices it at $30 / $200. Cursor Teams includes “usage analytics.”
• Uncrowded: fail. YC incumbent + OSS clone + vendor dashboards.
• 60s: fail without a Mac app / GitHub.
• 1-VPS: possible for a dashboard; the valuable bit is the desktop sidecar SuperPenguin already ships.
• Legal: fine if no prompt bodies (SuperPenguin’s default).

Alt 2 — MCP blast / scan (CodeFlowOps shape)

KILL (no monthly WTP)

• Pain: mild, at paste-time. Claude defaults MCP to always_ask because this fear is already in the client.
• WTP: ShieldMCP $49 one-time, 5 PH upvotes. CodeFlowOps is free and waitlisting “continuous verification” — they do not have a converting sub either.
• Uncrowded: fail.
• 60s: pass.
• 1-VPS: pass (read-only).
• Legal: pass if we never call tools and never hold tokens.

Alt 3 — English permit + receipt (Slip / Hallpass)

KILL (no demand, crowded mechanism)

Covered in §5. HN 2 and 9 points. Native policy in Cursor/Claude. Permit.io Community free.

Alt 4 — Auth / billing / MoR for weekend AI products (tiun / Creem shape)

KILL (legal + crowded)

• Pain: the hottest PH job this month. “Agent shipped the app; you spent three weeks on Stripe.”
• WTP: they already pay Stripe; MoRs take 3–6% of GMV.
• Uncrowded: fail.
• 60s: their aha, not ours.
• 1-VPS: fail (tax, payouts, chargebacks, KYC).
• Legal: fail. Money transmission / MoR. BRIEF constraint 3.

Alt 5 — Put the coding agent in Slack / Teams (Switch shape)

KILL (crowded, OSS, incumbents)

• Pain: “the team lives in Slack, the agent lives in a terminal.”
• WTP: unproven as a $29 sub; Switch launched OSS/self-host (PH #4).
• Uncrowded: fail. Claude, Cursor, ChatGPT already have Slack surfaces.
• 60s: maybe a demo room.
• 1-VPS: possible; WebSocket/Slack ops is a support furnace.
• Legal: fine if we are not logging into third-party sites as the user.

Alt 6 — Quantum classroom

KILL (IBM free + real hardware)

Covered in §2.5.

Alt 7 — Website visitor → booked meeting (Naoma)

KILL (crowded, not 1-VPS)

Video demo agents, avatars, CRM routing. Sales-led. PH traction is not our motion.

No alternative survives. The jobs with card evidence are owned by model labs, Cursor, Lindy, Langfuse, SuperPenguin, Stripe/MoRs, or IBM. The jobs that look “open” have HN/PH evidence of indifference.

7. If a job survives

None does. The template is filled in the negative so it cannot be hand-waved later.

8. What this means operationally

1. Do not cut nginx. Do not replace cryptobook.space. Incomplete files under /root/cryptobook-next/app/ are a prototype of a product without a buyer.
2. Do not “patch Slip again.” Another compiler/receipt/card remix is the same failed process.
3. Do not fall back to quantum so the homepage can say a 2026 word.
4. Waiting is cheaper than a useless product. The VPS already hosts paying properties (tokensmarketcap, VBL). Burning founder hours on a $0-WTP sidecar is the real COGS.
5. If a later pass finds a job, it must arrive with invoices, a named incumbent, and a reason that incumbent will not ship it this quarter — not with a survival spreadsheet.

Final line

NO-GO: no idea has enough demand evidence to build on this VPS in Sep 2026

Sources actually opened (18 September 2026)

1. cursor.com/pricing — Hobby free, Individual $20/mo, Teams $40/user/mo, Enterprise custom.
2. claude.com/pricing — Pro $20/mo ($17 annual), Max from $100, Team $20–25 / $100–125 per seat, Claude Code on paid plans.
3. lindy.ai/pricing — Plus $29.99, Pro $99.99, Max $199.99 per user/month.
4. langfuse.com/pricing — Hobby $0, Core $29, Pro $199, Enterprise $2,499; MIT self-host.
5. langchain.com/pricing — Developer $0, Plus $39/seat, Enterprise custom; LCU $1.50, LSU $1.00.
6. langchain.com/state-of-agent-engineering — 12 Jun 2026 survey, 1,340 responses.
7. superpenguin.ai — Free / Growth $30 / Pro $200; not a proxy.
8. www.permit.io/pricing — Community free forever; Enterprise contact us.
9. Okta Agent SSO press, 24 Aug 2026 — $0 extra on core SSO.
10. Helicone × Mintlify, 3 Mar 2026 — maintenance mode; 14.2T tokens, 16k orgs.
11. ibm.com/quantum/pricing — Open free, $96/min PayGo, Flex $72/min, Premium $48/min.
12. IBM Classroom Accounts, 28 May 2026 — no credit card, real QPUs.
13. Menlo / GlobeNewswire, 16 Sep 2026 — $40B consumer spend, 55% pay, 41% tried an agent.
14. Product Hunt monthly, Sep 2026 — tiun, Mastra, Kilo, Switch, GPT-6, Monid, Naoma, Voiskey, Anysite.
15. ShieldMCP on Product Hunt — $49 one-time PDF, 5 upvotes, 0 reviews.
16. codeflowops.com — free MCP scan, CLI, waitlist for continuous.
17. HN: Permit MCP Gateway — 2 points, 1 comment.
18. HN: cMCP signed receipts — 9 points; AI-generated comment flagged.
19. HN: l6e agent budget MCP — 6 points; $1,263 Cursor bill.

Internal corpus read (not treated as gospel): decision/PM-NOTE.md, decision/DECISION.md, research/*.md, challenges/*.md, ideas/*.md, BRIEF.md.