Product & platform research · propose-only · 2026-07-04

Beyond the Wrapper — Product & Platform Research

A follow-up to the TrustMRR climb — the research this one builds on. That page asked "how do we reach the top 25." This one asks the harder question: what do we build on Core + Model so a customer either makes money or loses a real pain — and no beginner vibe-coder can clone it? The brief was deliberately wide: explore both B2B and B2C, all four moat directions, greenfield (best idea wins), inspired by real YC / TrustMRR / indie-hacker startups. This is desk research to inform strategy — it feeds the Strategy doc + Backlog, it does not decide them.

TL;DR

  1. The spine is NOT the wrapper — the products are. Sam's instinct is right that humaniz / trendfy / speedback are thin. But Core (auth + billing + email) plus the Model gateway — "any model, pay once, no subscription, EU / on-device" — is exactly the infrastructure that indie winners and YC both name as the real moat. The fix isn't "be less of a wrapper" everywhere; it's to add one moat type to each product: proprietary compounding data, workflow/integration depth, or privacy/on-device.
  2. First-dollar and "bigger than a weekend project" are the same move when we ship done-for-you outcomes with distribution baked into the build. Near-term: the headshot (foto-service) via the free-tool-as-SEO front door (the Betterpic pattern), pay-once with a hard cap. Platform bet: turn Core + Model into a monetized platform other agents call — an MCP gateway — the thing no beginner ships.
  3. Distribution is the actual constraint, so pick products where engineering IS the distribution. Free tools, programmatic SEO pages, and marketplace / MCP listings are things the dark factory builds, not things Sam has to hustle daily. A great product no one cares about — or one people think they can just ask GPT for — is a loss.

1 · What we already have — the unfair advantages

Before shopping for products, inventory the assets. Most indie builders and MCP hackers are missing the exact things we already run in production — that is where a moat starts.
13
Core endpoints, LIVE at core.oll.am — passwordless auth · Stripe · email · health. ~300 lines, frozen.
4
Model providers behind one switch at model.oll.am — groq · claude · ollama (self-hosted) · mock.
~50–200
Lines to add a whole new product — routes + a service module + an OpenAPI slice. Everything else is reused.
EU
Frankfurt hosting → an honest "your data never leaves the EU, or your device" claim.

Core — magic-link auth (72h HS256 identity-only JWT), Stripe billing (checkout / webhook / verify / portal, one-time AND subscription, multi-product via an env price map, idempotent webhooks), Resend transactional email, health probes. Most indie and MCP builders lack a billing spine entirely — we amortize ours across every product we ship.

Model gateway (oll-model) — a provider switch across groq / claude / ollama (self-hosted, on our VPS) / mock. Stateless, key-custody, typed error classification, token accounting. Today it is a relay — no streaming, tool-use, or routing yet. That is precisely the upgrade surface Section 3 mines.

The dark factory — contract-first OpenAPI → DTOs → client, a Schemathesis drift gate, a keyless CI integration compose, per-service Coolify deploys, auto-merge on green. This is why we can co-optimize product AND distribution overnight: a new product is mostly reuse.

Guardrail on the privacy claim. NEVER say "Swiss-hosted" or claim CH data-residency — the VPS is in Frankfurt. The true, sellable claim is "EU data residency + optional on-device." "Swiss-made" is irrelevant; keep only the claims that are code-true.

Existing POCs as raw material (greenfield — treat these as optional beachheads, not commitments): foto-service (a real per-order LoRA headshot pipeline, guest checkout, 45 tests), oll-write (13 text-ops endpoints, SSE), ollwrite (a productized editor, not deployed), specview (spec + SQL), humaniz (frontend-only on Core + write), trendfy, speedback.

2 · What the outside world says — three converging signals

Indie hackers, YC, and the infra research literature disagree on almost everything except where the moat now lives. That agreement is the finding.

Indie hackers & TrustMRR

Money is in (a) bounded consumer deliverables whose output makes the buyer money or look good — Photo AI ~$132K MRR, Pikzels (YouTube thumbnails) ~$25K, Betterpic (free-headshot SEO → paid) — or (b) legible-ROI B2B (Churnkey recovers churned revenue). Pure wrappers die: Jasper fell ~$120M → $35–55M ARR when ChatGPT caught up.

YC (2024–26)

>50% of the Spring-2025 batch builds agentic AI. The winning wedge: own ONE narrow, painful, auditable job end-to-end and become its system of record (Cubic, Everest, Hypercubic, Eloquent). YC's stated moat is NOT a proprietary model — it's evals + integration depth. Two live RFS describe our spine: "SaaS Challengers" and "Software for Agents."

The moat moved

Out of the model call, into the infra: multi-model routing, structured extraction, MCP, RAG, evals, on-device / EU. MCP is now a cross-vendor standard (donated to the Linux Foundation's Agentic AI Foundation, Dec 2025; 10,000+ public servers, <5% monetized — an underpriced tollbooth). EU AI-Act sovereignty is real regulation-driven WTP (Gartner: sovereignty inquiries +305% H1 2025).

The margin trap (why pricing has to be deliberate): AI-SaaS gross margin runs 50–60%, not the classic 80–90% — and 78% of buyers hit surprise usage bills. So pay-once ONLY works on bounded deliverables or with an on-device offset. Winners distribute via build-in-public (Pieter Levels' 10-year X audience — Photo AI did $5.4K in week one), free-tool SEO, and marketplaces — not paid ads.

CompanyWhat it doesTech edgeDistributionTransferable lesson
Photo AIAI headshots / photoshootsPer-user tuned image modelBuild-in-public (Levels' audience)Bounded deliverable + audience = fast $
BetterpicAI headshotsQuality + turnaroundFree headshot generator → SEO → paidThe free tool IS the funnel
ChurnkeyChurn / retention recoveryPayment + dunning integrationB2B outbound, legible ROISell recovered revenue, not features
PikzelsYouTube thumbnailsCreator-tuned generationCreator communitiesNarrow ICP + output that earns them money
HarveyLegal AIDomain data + evals moatEnterprise / design partnersVerticalize into jargon + compliance
Cubic (YC)AI code reviewIntegration + eval depthDev channels / GitHubOwn one auditable job end-to-end
Metorial (YC)MCP runtimeHosted, metered MCP infraMCP directories / devThe tollbooth layer is fundable
Marc LouPortfolio of micro-SaaSShared boilerplate spineOne audience across many productsA shared spine + audience > one big app

3 · The centerpiece — 10 platform extensions

Each row: an infra capability we could add to Core + Model → the features it unlocks → a concrete product → real comparableseffort. Every one rides the spine we already run.
#ExtensionFeatures it unlocksProduct to buildComparablesEffort
1 MCP gateway Our tools as auth'd + metered MCP servers; tool-calling into oll.am; billed via Core "oll.am tools" — headshot / writing / extraction callable inside ChatGPT / Claude / Cursor Smithery, MCPize, Zapier MCP Med
2 Structured extraction Schema-locked JSON from any input; forms / receipts / contracts → typed data "receipt / invoice → accounting CSV" pay-per-batch Reducto, LlamaExtract, Klippa Low
3 Document / vision ingestion Agentic OCR / VLM for messy PDFs & scans; RAG-ready ingestion; contract parsing "stack of PDFs → clean structured export" Reducto, LlamaParse, Docling Med
4 Hybrid RAG-as-a-service Vector + keyword + rerank + chunk-ACL over a user's corpus; "chat your docs" with citations + access control EU-hosted "chat your legal / medical files" Vectara, Ragie, Morphik Med–High
5 On-device / EU-sovereign tier Local (Ollama) + EU-hosted routing as DEFAULT; a privacy-guaranteed variant of every product; regulated-industry sales "private writing tool — runs on your machine / EU-only" Mistral Le Chat, Ollama apps Med
6 Eval + guardrail harness Offline regression evals + inline guardrails; a reliability SLA; brand-safety Internal quality gate that becomes a selling point ("we test every output") Braintrust, DeepEval, Guardrails AI Med
7 Agent / orchestration runtime Stateful checkpointed multi-step workflows; human-approval steps "research agent → cited report" LangGraph, CrewAI, OpenAI Agents SDK Med–High
8 Voice layer (STT / TTS / real-time) Streaming transcription + speech; dictation; phone agents; voice-notes → structured output "voice-note → structured spec / email" (fits braindump-to-spec) Deepgram Voice Agent, Vapi, ElevenLabs Med–High
9 LoRA fine-tune + distillation Cheap per-customer / style tuning; brand-voice models; a style prompts can't reach; cheaper inference "writing tool tuned to YOUR voice" (a data flywheel) Together FT, Replicate LoRA, Predibase Med
10 Browser-automation worker DOM + vision agent for API-less sites; scrape → structure; auto-fill "auto-fill / competitor-watch" agent Browserbase + Stagehand, Browser Use, Playwright MCP Med
Highest leverage-per-effort FOR US = #1 MCP gateway · #2 structured extraction · #5 EU / on-device tier. Each rides the existing spine, each has a 2026 willingness-to-pay tailwind, each is an overnight CI slice. The rule that falls out: BUILD the layers tied to customer-specific state (integrations, documents, a tuned model) — that's where a moat compounds — and CONSUME the commodity layers (routing engines, browser infra, base models).

4 · Products we could actually ship

Greenfield candidates, not commitments. Each card: audience · why-pay · moat type · distribution baked in · comparable · overnight-feasibility.

Headshot Studio (foto) — B2C

Why pay: pay-once bounded deliverable; the output makes you look hireable. Moat: done-for-you + a per-order LoRA. Distribution baked in: a FREE watermarked generator as the SEO front door + programmatic "AI headshots for {profession}" pages. Comparable: Photo AI / Betterpic. Feasibility: already scaffolded — the near-term first dollar.

Ledger Snap — B2B / prosumer

Why pay: "receipts / invoices → accounting-ready CSV / Xero" removes bookkeeping hours (legible ROI). Moat: structured extraction + integration. Distribution: an MCP server + a Chrome / marketplace listing. Comparable: Klippa / Reducto.

Vault Chat (EU-private RAG) — B2B regulated

Why pay: "chat your documents — data never leaves the EU or your machine"; WTP driven by the EU AI Act. Moat: hybrid RAG + EU / on-device. Distribution: LinkedIn design partners in one vertical. Comparable: Vectara / Morphik.

oll.am Tools — the MCP gateway (the platform bet) — B2B / prosumer

Why pay: our tools callable and billed inside ChatGPT / Claude / Cursor. Moat: distribution-as-moat + the billing spine others lack. Distribution: MCP directories (mcp.so / Smithery) where <5% are monetized. Comparable: Zapier MCP / Smithery.

Voicespec — prosumer

Why pay: "voice-note → structured spec / email / CRM entry" removes the blank-page + transcription pain. Moat: voice + structured extraction + our braindump-to-spec engine. Distribution: a free tool + build-in-public. Comparable: Superwhisper / AudioPen.

These are candidates, not commitments. Pick by the co-optimization test in Section 7 — not by which is most fun to build.

5 · Product-market fit & the "just ask ChatGPT" test

Validate on behaviour, not opinion. Run the Sean Ellis 40% test — "how would you feel if you could no longer use this? very disappointed / somewhat / not" — on the first ~40 users, and gate it on week-4 retention, because AI novelty spikes then churns (a launch spike is not PMF). And sell before you build: a Stripe link + a landing + a waitlist deposit proves willingness-to-pay before a line of product code. The CHF 1 / CHF 9 pattern already live on oll.am is this test.

The "just ask ChatGPT" objection IS the PMF question. There are seven honest answers to it; our realistic three are in bold:

Answer to "why not just ask ChatGPT?"What it meansOurs?
Done-for-you outcomeA finished result, not a capability✅ headshot, extraction
VerticalizationA narrow ICP's jargon / compliance / integrations✅ Ledger Snap, Vault Chat
Workflow integrationLives inside the tool you already usepartial (MCP)
Proprietary-data flywheelGets better with each customer's datalater (LoRA voice)
DistributionBe where the user already is✅ free tools, MCP dirs
Guarantees / complianceAn SLA, EU residency, brand-safetypartial (EU tier)
Purpose-built UX on a repeated taskFaster than re-prompting each timepartial

The two-sentence case study: Jasper stayed a wrapper and got run over when ChatGPT caught up. Cursor became the workflow — the place the work happens — and is valued around $29B. The difference is not model quality; it's whether you own the outcome, the vertical, or the distribution.

6 · Distribution, ranked FOR US — and pricing

Sam is a strong engineer with weak distribution. So the ranking is not "what works in general" — it's "what an engineer can build instead of hustle." Bar = fit for us.
Free-tool / PLG
1 · best
Programmatic SEO
2
Marketplaces / MCP
3
Communities / Reddit
4
Build-in-public
5
Cold outbound (B2B)
6
Integrations / partners
7
Paid ads
8 · last
  1. Free-tool-as-SEO / PLG — engineering IS distribution. Betterpic drove ~8k clicks/mo → paid off a free generator.
  2. Programmatic SEO — Dynamic Mockups grew traffic +220% off generated pages ("AI headshots for {profession}").
  3. Marketplaces — MCP / GPT / Chrome / Notion. A Zillow→CRM extension hit $4.2K MRR in 4 months off a listing.
  4. Communities / Reddit — one honest, specific thread, indexed for months. Problem-first, never a pitch.
  5. Build-in-public (X / LinkedIn) — the oll.am build-log already IS this; it just needs an audience-facing channel.
  6. Cold outbound — B2B first-10 customers only; does not scale, but proves the pain.
  7. Integrations / partnerships — be the "AI action layer" on someone else's system of record.
  8. Paid ads — LAST. Thin AI-SaaS margins punish CAC; only after organic unit economics are proven.

Pricing — the deliberate part

Pay-once for BOUNDED B2C deliverables with a hard usage cap — never unbounded pay-once + LLM, it bleeds margin (recall the 50–60% gross-margin reality). Hybrid credits / outcome-based for B2B and any unbounded usage. Cautionary tale: Cursor's opaque credit switch burned users into a public apology — pricing changes are trust events. Price so the buyer clearly makes money or loses a pain: B2B tolerates $75–497/mo when ROI is legible; B2C wants single-digit pay-once and hates surprise bills.

7 · The co-optimization test & recommendation

Propose-only. This is the gate + the three moves that fall out of it — it feeds the Strategy doc and the Backlog; it does not edit the plan.
The co-optimization rule. Before we build, a candidate must pass BOTH: (a) a moat — one of compounding data · workflow-integration · EU / on-device — so it isn't a clonable wrapper; AND (b) a distribution channel baked into the build — free-tool SEO, programmatic pages, or an MCP / marketplace listing — so we're not left hustling. Products that fail either test are exactly how we got thin GPT wrappers.
  1. NEAR-TERM — the first stranger dollar. Ship the headshot (foto) with the Betterpic free-tool-as-SEO front door + pay-once-with-cap. Proven pattern, already scaffolded, distribution baked in. It passes both halves of the test.
  2. THE PLATFORM BET (bigger than a weekend). Build the MCP gateway — expose our tools as auth'd, metered MCP servers. It's distribution AND defensibility in one move, it answers a live YC RFS ("Software for Agents"), and it's the "no beginner vibe-coder does this" play — because it rides the billing spine almost no MCP builder has.
  3. THE MOAT DISCIPLINE. Adopt the co-optimization test as a standing Backlog gate, and add the EU / on-device tier + structured-extraction as the first two shared extensions — so every future product inherits a moat and an honest privacy claim by default.

Honest caveat (house style). This is desk research synthesizing public build-in-public numbers — directional, not audited — plus our own code. The next step is to turn it into a scored Backlog, not to build all ten. Nothing here changes the plan of record: the Backlog and Strategy decide; this informs them.

Indie hackers / moats hatchworks (wrapper moats) · Menlo & Vendep (vertical / workflow moat) · Indie Hackers · Flowjam (27 examples) · founderpath / Betterpic (free tools) · omnius (programmatic SEO)
YC / market ycombinator.com/rfs · CB Insights (YC Spring-25 agentic) · Forbes (YC Fall-25) · YC Library (7 moats) · BVP State of AI 2025
Infra / MCP / routing OpenRouter routing docs · Anthropic (MCP / AAIF donation) · WorkOS (MCP-2026) · aakashg (AI pricing)
EU / PMF Lyceum & Impetora (EU data residency) · pmfsurvey.com (Sean Ellis 40% test)

Follows The TrustMRR climb. Propose-only · 2026-07-04 · reference links are honest and directional, public build-in-public figures are not audited.