Finance challenge — CryptoBook pivot

Date: 18 September 2026
Role: adversarial finance. No hopium.
Bar: $5,000 MRR by month 12, one Ubuntu VPS, content + SEO + product-led growth only. No enterprise sales, no GPU cluster, no prepaid LLM wallet.
Anchor price: Langfuse Cloud Core $29/mo (100k units, 90-day data, unlimited users). Pro $199. Hobby $0. Self-host MIT $0.
Ideas directory: empty after a 3-minute poll. This file therefore scores the four candidate directions in BRIEF.md (A–D) plus one patched SKU that exists only because A–D fail the bar as written.

Method (the rules I will not bend)

1. Sunk VPS is not a business. Host is already paid (~Contabo-class, 94 GB RAM, 678 GB disk, fair-use unmetered traffic, throttle risk if we become a token proxy). Incremental compute ≈ $0. That does not mean the idea is cheap. Support hours, Stripe, email, and any LLM we run are the real COGS.
2. $5k MRR is a customer-count problem, not a TAM-slide problem. At $29 you need 173 paying accounts. At $19, 263. At $49, 102. At $99, 51. Bain’s “$100B US agentic SaaS” (May 2026) and MarketsandMarkets’ $19B agentic number are irrelevant. We will not touch 0.001% of that from cryptobook.space.
3. Distribution tax is real. cryptobook.com is a social-media leftover. SEO for “MCP gateway / agent permit / quantum classroom” from a brand named CryptoBook converts like a casino site selling CI tooling. I apply a 30–50% conversion haircut unless the product name on the page is not CryptoBook.
4. Free incumbents set WTP, not Langfuse. Cloudflare AI Gateway spend limits shipped 5 Jun 2026 (Workers Free: 100k logs). PolicyLayer hosted MCP gateway is free-to-start; Intercept is a local OSS binary. LiteLLM hard-caps spend. IBM Quantum Composer + Qiskit textbook are free. agentreceipts.ai is Apache OSS. If the paid SKU is “hosted convenience of a free tool,” conversion is 1–2%, not 5%.
5. Benchmarks I will actually use (2026, not LinkedIn): - Freemium → paid median ~4.5% (GrowthSpree 2026); 2–8% range. Dev-tools with a strong OSS alternative: use 2–3.5%. - Visitor → signup, ungated playground: 3–6%. New unknown domain: 2–4%. - Developer-tools monthly logo churn: 2.7–3.8% if embedded in a pipeline; 4–6% if web-only or set-and-forget; 6–10% under $1k ACV monthly billing (ChurnCost 2026). - Solo-founder support: 3–5 h/week under $5k MRR is survivable; >10 h/week is a hidden COGS that eats the founder. Proxy/auth products run hot.
6. Stretch is allowed only if the operational lever is named. “It could go viral” is not a lever. A high-intent keyword + a 60-second playground + a paywall tied to a felt cash event is a lever.

Gross margin on this VPS for a non-LLM-hot-path SaaS is 85–92% after Stripe. Nobody dies on COGS here. They die on WTP, churn, and not enough paying humans.

Shared 12-month cost sheet (all ideas, this VPS)

Rule: any idea whose hot path is an LLM we pay for, or whose support is “debug my agent’s SSE proxy,” has a founder-hour problem before it has a cash problem.

Verdict board

A — Permit gateway for agent tools/MCP

What it is (from brief): English rules, allow/deny/ask, spend cap on the user’s own keys, signed shareable receipts. One URL drop-in. Not enterprise GRC.

Unit economics (as specified, $29 to “match Langfuse”)

COGS. Fine. Deterministic policy engine. Do not proxy full LLM bodies. Receipt = hash + decision + spend delta. English→policy: compile on signup with gpt-4o-mini or a local 8B on this box. Year-1 LLM bill if we are disciplined: <$50. Disk: hashes. Bandwidth: MCP tools/call JSON, not tokens.

WTP vs Langfuse $29. Wrong comparison, and buyers will make it anyway if we put $29 on the page. Langfuse sells traces to ML teams. This sells a kill-switch to the human who pays the API bill. Cloudflare already does LLM spend limits for $0. PolicyLayer already does hosted MCP allow/deny/ask, free to start, plus OSS Intercept. Docker / agentgateway / LiteLLM / Lunar MCPX / MintMCP exist. MintMCP is quote-only enterprise — good, we should not go there — but it means “MCP gateway” as a category is already a sales-led pile-up.

Honest WTP:

$29 is the worst possible price. It is Langfuse’s number, above indie insurance WTP, below team WTP, and it invites a feature-matrix death vs PolicyLayer.

Conversion. Playground can be <60s: type a rule, get a URL. That part is financially real. Free→paid will still be ~2.5–3.5% if the free tier is a working gateway (PolicyLayer already gave that away). Paywall “more calls” is a usage meter against a free hosted competitor. Paywall “signed shareable receipts + 90-day inbox” is a better wedge.

Churn. Set-and-forget cap = 6–8%/mo. User sets $5/day, survives a week, cancels. Without a daily deny-log / spend inbox, NRR < 80%. With a weekly email of “your agent was denied 14 times, saved $62,” churn can sit at 4.5%/mo. Still ugly on monthly billing.

Support. This is the silent killer. MCP streamable HTTP + Cursor vs Claude Desktop vs ChatGPT connectors + OAuth + SSE is a 20–40 minute ticket. Use 0.4–0.8 tickets per paying user / month. At 173 payers (the $29 × $5k headcount) that is 70–140 tickets/mo ≈ 25–70 founder hours. That is not a SaaS. That is a consultancy with a login page. Docs-gate free users or die.

12-month P&L (as specified, CryptoBook-branded, $29, generic “gateway”)

Assumptions: visitor→signup 2.5% (brand tax), free→paid 2.5%, churn 6%/mo, ARPU $29, SEO ramp to 15k visits/mo by M12.

Does not clear $5k. Stretch to $5k requires 173 payers, which at 2.5% conversion needs ~7,000 living free users, which at 2.5% visitor→signup needs a traffic run-rate this domain will not have in 12 months for the keyword “MCP gateway.”

Verdict: FIX

Keep the wedge. Kill the positioning. Mandatory finance changes or this is a KILL on re-review:

1. Do not price at $29. SKUs: $19 indie / $49 studio (5 seats). Blended ARPU target $31–$39. Never put Langfuse’s number on the page.
2. Do not sell “MCP gateway.” PolicyLayer and Cloudflare already occupy that sentence. Sell the receipt the human forwards and the inbox of denials/spend. Gateway is the pipe, not the SKU.
3. Do not proxy LLM tokens. MCP tool calls + optional BYOK spend counter only. Token proxy makes us Helicone-at-$0-margin and trips bandwidth.
4. Do not store traces. Hashes + decision + $ delta. Disk and GDPR/AI-Act logging theater stay cheap.
5. Paywall = shareable receipt + 90-day inbox + ask-me-first, not “more events.” Events are what Langfuse meters; we will lose that race.
6. Product name ≠ CryptoBook. Domain can stay. The H1 cannot say CryptoBook. Brand tax otherwise eats the conversion the playground is supposed to create.
7. Free support is docs-only. Paying gets email. Otherwise founder-hour COGS exceeds cash COGS at month 8.
8. No SOC2, no “AI Act compliant,” no SSO in v1. Those are MintMCP’s $quote motion. They also pull us toward EU high-risk claims we are forbidden to make.

With that list, this collapses into idea F below.

B — Repo agent-permit.yaml runtime (Dependabot-like)

What it is: YAML in the repo, hosted enforcement, PR-viral like Dependabot.

Why the finance story is fan fiction

Dependabot’s distribution is GitHub. Ours is a README. GitHub Marketplace apps without a pre-existing audience realistically do 300–1,000 installs in year 1. Top-quartile might see 2,000 if an HN post hits. That is the entire funnel.

Stretch $7.9k requires both a viral GitHub App and $99 WTP for a YAML file whose useful core will be MIT-cloned in a weekend. I will not underwrite that.

WTP vs Langfuse $29: worse. This is closer to a linter. People pay $0 for .eslintrc. They will pay $0 for agent-permit.yaml and run it in Actions. Hosted enforcement is only valuable if CI cannot see secrets or if you need a public URL. That is a $9–$19/mo feature for a handful of orgs.

COGS: excellent (no LLM, no proxy). Support: YAML schema + CI + GitHub App permissions is still 30-minute tickets. Churn: high. Once the YAML exists, the hosted part looks optional.

Viral loop math: Dependabot-like PRs only fire if thousands of repos already have the file. Chicken, egg, 18–36 months. We have 12.

Verdict: KILL

A YAML spec can be a file format under F, not a company. Do not build a GitHub App as the product. There is no path to 102 paying orgs on content+SEO in 12 months that does not assume a miracle launch.

C — Freelance / client agent receipts

What it is: prove to a client that an agent did the work.

Ceiling test (this is sufficient to kill)

Who pays?

• Freelancer: already delivers a git diff, a Loom, and an invoice. Cryptographic receipts do not get them paid faster. WTP $0–$9, mostly $0.
• Client: does not know what an Agent Receipt is. Will not install a verifier. Chicken-and-egg.
• Protocol layer: agentreceipts.ai / obsigna is free OSS (Ed25519, W3C VC, MCP proxy). Dropwatch-class MCP already meters $0.01 USDC/call. We would be a hosted UI on top of a free spec.

TAM of freelancers who (1) bill with agents, (2) have clients who distrust them enough to want a hash chain, (3) will pay monthly: a few thousand globally if we are generous. At 100% awareness, 2% paid, $9 ARPU:

Ceiling MRR ≈ $900. That is the happy ceiling. Base case is 10–30 paying hobbyists, $90–$270 MRR.

Conversion from free→paid: worse than generic freemium because the value accrues to the other party (the client) who is not the buyer. Classic two-sided trap. We are not allowed to become a payments rail to solve it (no money transmission).

Churn: after the gig ends, they cancel. Project-based usage on a monthly sub is poison. Per-receipt billing at $0.01 is a protocol, not a $5k SaaS.

Support: low volume, so support is fine. Irrelevant.

Vs Langfuse $29: no overlap, and that is not a compliment. Langfuse has a daily user. This has a user on invoice day.

Verdict: KILL

Too narrow, WTP is a rounding error, OSS protocol already exists. A shareable receipt is a feature of F, not a customer segment.

D — Quantum circuit classroom

What it is: educational circuit UI, probably with an LLM tutor, on this VPS.

Demand is not the same as dollars

IBM Quantum Composer is free. Qiskit textbook is free. Quirk is free. qBraid QUEST (Jul 2026) is paying universities up to $5,000 in QPU credits per course. SpinQ sells hardware to institutions. Q-CTRL Black Opal is a funded education product. A black-and-pink circuit toy on cryptobook.space is not a wedge. It is a tutorial.

WTP:

• Students: $0. EdTech freemium→paid is the worst in the First Page Sage 2026 table (2.1%).
• Hobbyists: $9 if they are lost. They are not lost. They have IBM.
• Universities: $2k–$20k site licenses via procurement. That is enterprise sales, 9–18 month cycles, forbidden by the bar.

COGS split — this is the knife:

A 30-qubit statevector is 16 GB; 32-qubit is 64 GB. This box can simulate a classroom circuit. That is not a product. It is a demo. Fault-tolerant IBM Starling is ~2029. Application-layer quantum SaaS without hardware is the brief’s own likely kill, and the finance agrees.

Churn: course-shaped. Semesters end. 8–12%/mo if anyone paid.

Support: “why is my GHZ circuit wrong?” × infinite students. Unbounded.

12-month PLG: SEO for “quantum circuit simulator” is owned by IBM. M12 paying: 0–40. MRR $0–$400. Stretch $1.2k only if we sneak into department budgets — which is a sales motion we are not doing.

Verdict: KILL

Do not ship a quantum product on this pivot. Not as a classroom, not as “QCaaS-lite,” not as a blog with a Bloch sphere. Application-layer quantum is underfunded for a reason: the buyer is an institution and the hardware is someone else’s.

Automatic kills (if an idea file appears later)

Score KILL without a full model:

• Anything whose homepage is traces/spans/generations. That is Langfuse Hobby $0 / Core $29 / MIT self-host.
• Anything that resells model tokens or holds a prepaid wallet. Legal landmine + COGS we cannot control.
• Anything that needs SSO/SCIM/SOC2 to close. That is MintMCP/Okta quote-ware. Time-to-$5k > 12 months.
• Anything that needs more than this VPS (GPU cluster, Kafka, ClickHouse replica, multi-region).
• Anything whose WTP exists only after a security questionnaire.
• Anything that claims “AI Act compliance.” We are not an auditor; also a lawsuit.
• Crypto tokens, agent wallets that hold client funds, CFAA-flavored “the agent logs into sites for you.”

F — SURVIVE: Permit Inbox (the patched SKU)

This is A after the FIX list, with C’s receipt as a feature and B’s YAML as an export. It is not a new category slide. It is the only SKU whose 12-month arithmetic can hit $5k without lying.

What we sell (one sentence)

A kill-switch and daily spend/deny inbox for coding agents: English rules, hard allow/deny/ask on MCP tools, hard cap on the user’s own LLM keys, signed shareable receipts the human can forward to a boss or client. One HTTPS URL. Value in the playground in <60 seconds. We never resell tokens.

Not: an MCP control plane, a Langfuse, a GRC platform, a freelance marketplace, a quantum lab.

Why WTP exists (cash, not vibes)

The buyer already pays Anthropic/OpenAI/Cursor $20–$200/mo. A runaway loop is a $400 weekend — this is a real, public complaint class in 2025–2026. Insurance WTP is 5–15% of the insured monthly bill → $12–$19 indie, $39–$49 when 3–5 people share a deny-log. Cloudflare caps LLM $. We cap tool effects (bash, git push --force, email.send, payments MCP) and optional BYOK $ . PolicyLayer is the closest clone; we differentiate on English → policy in the playground and a receipt that looks like Stripe’s, not a log table.

If we cannot make a receipt a human actually forwards, this idea collapses back to A-as-specified and should be killed.

Pricing (do not “match Langfuse”)

Blended ARPU target: $31 base (60/40 indie/studio), $39 if studio mix hits 40%+5%. Annual = 10 months (not 12). Monthly default. Hiding churn with 50% annual discounts is hopium.

COGS at target ($5k MRR)

This VPS is enough through well past $5k if we do not store traces and do not proxy tokens. Storage/bandwidth become a problem around Langfuse-scale ingestion, which we will not have.

Conversion and churn (harsh)

• Visitor → signup: 4% (playground is the site). 2.5% if the H1 says CryptoBook.
• Free → paid: 3.5% base / 5% stretch. Stretch only if the paywall is the receipt they want to send the same day they got billed. If we meter “events,” use 2%.
• Churn: 4.5%/mo with weekly inbox; 7%/mo without. First 90 days are the graveyard.
• Implied: to hold 160 paying at 4.5% churn you need ~7–8 net new paying / month just to stand still, ~15–25 to climb.

12-month cohort (content + SEO + playground, no ads)

Traffic ramp (Permit-named, not CryptoBook-named; long-tail bill-shock + MCP allow/deny + “claude code spend limit”):

M1 0.8k → M3 2k → M6 7k → M9 13k → M12 20k visits/mo. Year-1 visits ≈ 105k. This is good execution, not median. Median with the CryptoBook H1 is half of that.

Base: 4% signup, 3.5% free→paid in 30 days, 4.5% monthly churn, $31 ARPU.

Misses $5k on base. I am not going to round it up.

Stretch (named levers, not vibes): 5% visitor→signup (playground is actually good), 5% free→paid (paywall = receipt-to-forward), $39 ARPU (studio mix), same churn.

128 × (5/4) × (5/3.5) ≈ 229 paying × $39 ≈ $8.9k MRR.

$5k sits between base and stretch. That is the only idea in this file for which that sentence is true. Levers, in order:

1. H1 / product name is Permit (or similar), not CryptoBook.
2. Paywall is the receipt + 90-day inbox, not quota.
3. Weekly “saved $X” email (churn).
4. One comparison page vs Cloudflare spend limits (“they cap tokens; we cap rm and git push --force”).
5. MCP directory / Claude Code hook snippet (distribution cheaper than SEO).
6. No LLM chat, no GitHub App as the product, no freelance marketplace.

If those six are not in the spec, downgrade F to FIX and do not ship.

Support load at $5k

~160–230 paying × 0.35 tickets/mo × 18 min ≈ 17–24 h/mo, plus docs-gated free noise ~8–12 h/mo. Total ~7–9 h/week. Solo-survivable. MCP proxy bugs will try to blow this. Ship a status page, a known-clients matrix (Cursor / Claude Code / Claude Desktop), and a copy-paste config. If ticket mix is >40% “my client won’t connect,” we mis-built the pipe.

What would make me kill F on re-review

• Pricing page leads with $29.
• Homepage says “MCP gateway” or “observability.”
• We store full prompts/tool results.
• We proxy completions.
• We hire a salesperson.
• We add an LLM tutor.
• M6 paying < 15 after a real launch (the playground is then not converting, and SEO will not save a dead activation loop).

Decision for PM

Build F (Permit Inbox). Do not build B, C, D. Do not build A as a generic gateway. If idea agents later dump a Langfuse clone or a quantum classroom into ideas/, they are already killed here.

Cash truth: on this VPS, COGS is not the puzzle. 173 humans at $29 is the puzzle, and $29 is the wrong number. 102 humans at $49 or 263 at $19, acquired from bill-shock and deny-events, with a receipt they forward, is the only 12-month path that is not a story.

Margin truth: 90% gross, founder support as the binding constraint. Stay off GPU, off traces, off enterprise questionnaires.

Brand truth: CryptoBook as a product name is a conversion fine we cannot afford. Keep the domain; bury the word.