3 AI wrapper teardowns: Jun 15–22 (one legal time bomb, one overnight sensation)

• CNVS's BYOK model (users bring their own Claude/Cursor/Codex subscriptions) is the compliant path forward now that Anthropic tightened commercial API terms in February 2026. If you're building an LLM wrapper, you need to read those terms before writing a line of code.
• Oli Ai demonstrates that Meta Pixel + 20% trial-to-paid conversion is a real playbook — but the TCPA exposure from Lowery v. OpenAI makes the AI voice calling niche one to approach with a lawyer on speed dial, not a weekend sprint.
• PhoneDiffusion's on-device positioning is an underexplored angle. Privacy-first AI image generation has no direct competitor running Stable Diffusion locally on iPhone. Low traction so far, but the moat is real.

Evidence tiers used in this article: [data] = publicly disclosed figure with a verifiable source; [editor's view] = the author's inference from available evidence.

Head-to-head replication scores

Oli Ai — AI voice appointment setter, $1,343 MRR

• Technical lift: ★★☆☆☆ — Twilio + voice synthesis API + LLM + Stripe. All public APIs, well-documented. The hard part is voice conversation quality and interruption handling, not the plumbing.
• Information edge: ★★☆☆☆ — No proprietary moat. A tuned Meta Pixel audience is reproducible with ad spend.
• Capital needed: ★★☆☆☆ — Meta ads at $13 CPA require real budget to build trial volume.
• Legal risk: ★★★★★ — This is the full stop. The FCC's February 2024 Declaratory Ruling classified AI-generated voices as "artificial or prerecorded voice" under TCPA, requiring prior express written consent for any telemarketing call. 4 In January 2026, a federal lawsuit named both Twilio and OpenAI as co-defendants for TCPA violations committed by a third-party telemarketer using their infrastructure. 4 TCPA attorney John Henson called it "the compliance bomb of 2025 — with reverberations that will shake the AI voice industry throughout 2026." There is no public evidence Oli Ai has implemented consent management, DNC scrubbing, or mandatory AI disclosure. Building a clone without these is building on quicksand.

• First concrete step: Before writing a line of code, spend $500 on a TCPA compliance consultation. Understand what prior express written consent requires, how DNC scrubbing works with Twilio, and what in-call disclosure language the FCC's pending NPRM will mandate. The tech is weekend-level; the compliance is not.
• First likely failure mode: You build the voice agent, the Meta funnel works, you hit 15 paying customers — and one recipient files a TCPA complaint. Statutory damages are $500–$1,500 per call. A class action is not a hypothetical.

CNVS — macOS AI agent canvas, $19,348 all-time

• Technical lift: ★★★☆☆ — Sub-300ms voice recognition, cross-agent memory, and a smooth visual canvas in a native Swift app is weeks of work for an experienced macOS developer.
• Information edge: ★★★☆☆ — No proprietary data, but daily user video calls and live stream feedback are generating a product roadmap a fresh clone won't have. [editor's view] Each week of shipping adds product decisions a fresh clone will have to reverse-engineer.
• Capital needed: ★☆☆☆☆ — YouTube streaming costs nothing; Swift dev requires only a Mac. No paid acquisition visible.
• Legal risk: ★☆☆☆☆ — BYOK is the compliant path under Anthropic's February 2026 terms. 10 No regulated industry exposure.

• First concrete step: Validate before writing Swift. Build a minimal Electron or Tauri app, let the user point their own Claude Code key at it, and render outputs on a zoomable canvas. Charge $49 one-time. If 10 people buy in a week, learn Swift.
• First likely failure mode: macOS native development is a skill most indie devs don't have. The app's tactile feel — Liquid Glass UI, live backgrounds, sub-300ms voice — is a native-only advantage. A web clone will feel slower, and that gap is half of CNVS's current word-of-mouth.

PhoneDiffusion — on-device AI art for iPhone, $321 MRR

• Technical lift: ★★★☆☆ — On-device Stable Diffusion via Core ML is well-documented, but optimizing across iPhone generations (A15 vs. A17) and managing model quantization takes real effort.
• Information edge: ★★★☆☆ — No proprietary dataset. No direct competitor identified that combines on-device Stable Diffusion with a polished iOS subscription model. [editor's view]
• Capital needed: ★★☆☆☆ — App Store listing is free; Core ML inference has no per-call compute cost. Main expense is dev time and optional paid UA.
• Legal risk: ★★☆☆☆ — Stability AI's license allows commercial use. On-device processing means no user data reaches the founders' servers, reducing GDPR/CCPA surface. The main exposure is user-generated infringing content — same as any image app.

• First concrete step: Download Apple's Core ML Stable Diffusion sample project from Apple's GitHub repos, run it on a physical device, and check generation speed on your target iPhone tier. That validation takes a day. If it's too slow, the moat just got bigger; if it's acceptable, you have a build path.
• First likely failure mode: App Store review is unpredictable for AI image generation apps. Apple has approved some and rejected others on content policy grounds — you could ship a complete app and sit in review for weeks.

This week's meta-signal: the compliance fault line is now a product decision

Sources

What to do tomorrow morning