Cornerstone Guide
A Candy-AI-style companion product, built with AI coding agents in about twelve hours: an uncensored brain, a character-consistent body, and a voice that sounds like an intimate conversation, not a robot assistant.
The article below is for you. The masterplan is for your AI. Hand it to Claude Code, Cursor, or any agent, and it takes it from there.
Start building
One click copies a prompt. Paste it into your agent and it opens the masterplan and reads it for you.
The stack
Venice runs an unrestricted roleplay model with 128K context, vision, function calling, and zero content filtering. It writes in-character, second-person dialogue, and it never narrates the scene or the user's side of the conversation.
WaveSpeed edits a fixed base image per companion with identity-preserving LoRAs. Her face and look stay stable in every picture, while themed styles are one trigger word away. The edit loop is what stops a companion from morphing.
SecretVoice delivers 29 premade intimate voices in 21 languages with sub-200ms streaming and realtime WebSocket. It sounds like an intimate conversation, not narration. No voice-actor rights to manage, ever.
The x/friend engine (MIT, by the SecretVoice team) already wires this trio together: rolling memory, durable facts, and send_image and send_audio tools. It absorbs the subtle 40% to 50% of the engine. You build the multitenancy, API, frontend, and billing on top.
The moat
Companion users do not pay for pictures alone. They pay to feel known, and voice is the shortcut to that feeling. A companion that occasionally sends a breathy voice message feels alive. Generic TTS sounds like an agent reading a script. SecretVoice is tuned for closeness.
Unit economics
No income promises. Just cost of goods, so you can price with margin on purpose.
| Interaction | Cost |
|---|---|
| Text turn (about 5k in, 250 out tokens) | about $0.003 |
| Voice message (about 300 characters) | about $0.025 |
| Image run | about $0.025 |
| A rich turn (text + voice + image) | about $0.05 to $0.08 |
A $29.99 Premium plan (unlimited text, 200 voice messages, 100 images) costs roughly $12 to $15 to serve at heavy usage and far less at typical usage. Even generous plans keep half or more of the margin. SecretVoice credits can auto-refill via USDC webhooks, so the voice leg never stalls.
The twelve-hour plan
Domain, email, all three API keys, base images, voice picks.
Set up the repo, install x/friend, and get the engine healthy.
Accounts, companions, memory. The multitenancy rule is the whole design.
Auth, companion CRUD, chat with streaming and tools, voice and image endpoints.
Swap the roleplay model, pin one voice per companion, tune the sampling.
Chat, voice player, image viewer, gallery, onboarding, paywall.
Billing and per-tenant metering with plan caps.
Age gate, disclosure, safety filters, zero-retention voice, deletion.
Deploy, then a smoke test: text to voice to image to restart to memory intact.
Guardrails
FAQ
No. AI coding agents (Claude Code, Cursor, opencode) generate the code from the masterplan. You review, run, and ship. The skill in 2026 is directing the agent, not writing every line.
At the masterplan stack, roughly $0.003 per text turn, $0.025 per voice message, and $0.025 per image run. Even a generous $29.99 Premium plan keeps most of its margin after cost of goods.
The masterplan splits the build into a twelve-hour timeline. With the x/friend engine already handling memory, compaction, and the image and voice tools, most of day one is API, frontend, and billing scaffolding.
For fictional characters and consensual adult use, yes, with guardrails: an age gate, AI disclosure, safety filters, zero-retention voice, and deletion rights. The masterplan covers each one.
SecretVoice ships 29 premade voices, 21 languages, and an ElevenLabs drop-in API. No voice-actor rights to license. Premade voices, license baked into the API terms.