Vibe coding credits incentive. The pricing model that fights the loop it sells.

Vibe coding only works if you iterate. Describe, watch the preview, change a sentence, watch the preview again. The loop is the product. Per-credit pricing on top of that loop creates a meter you cannot ignore: each iteration is a debit on a pool the platform owns. The rational user starts trying to one-shot the prompt so they do not burn the pool. One-shotting kills the loop. The platform is selling iteration and selling a tax on it at the same time.

It does not run a platform-owned credit pool. The first 6 anonymous turns share a hosted Claude key (ANON_TURN_LIMIT in src/app/api/chat/route.ts:23). After that the chat route at lines 217 to 228 tries the user's own Claude OAuth tokens first, and only falls back to the shared key if the user has not connected an account. When you connect Claude, every turn bills against your plan with no platform markup tier in the middle. When credits do run out, the classifier at route.ts:48-68 reads the upstream signal (billing_error, rate_limit, HTTP 402, HTTP 429) and surfaces it to the UI as credit_exhausted. That signal comes from Anthropic, not from a meter the product invented.

Because iteration is small. A good vibe coding turn changes one sentence in the prompt, watches the preview redraw, and decides whether the change helped. If every small turn debits a credit, the cost-per-decision climbs faster than the value of each decision. Users compensate by writing one large prompt that tries to anticipate every change ahead of time. The model is worse at large prompts than at small turns, so the output is worse, so the user blames the model. The actual cause is the pricing structure that pushed the user away from the loop the tool was supposed to enable.

Honestly, yes, with a different shape. ANON_TURN_LIMIT = 6 (route.ts:23) is the cost of having no signup at all. Without an account there is no identity to throttle, so the only abuse gate is the turn counter. Six is the smallest number where the visitor can close a real iteration loop (seed, refine, undo, retry). After that the route returns 403 sign_in_required and the UI shows a one-tap sign in. The honest framing is that the 6 turns are a free trial of the loop. Once you sign in and connect your Claude account, the incentive flips: every iteration bills against your Claude plan, which means the platform no longer profits when you iterate more.

It is a subscription, not a per-iteration meter. The distinction matters. A meter charges you per turn, which gives the platform an interest in your iteration count. A flat subscription does not. The Pro plan unlocks the workshop surfaces (the dashboard, multiple ongoing projects, model picker, longer sandbox lifespan), and the AI calls inside it run on either the hosted shared key or, preferably, your connected Claude account. The pricing page is intentionally not 'buy 100 credits.' It is 'buy access to the workshop, then connect a Claude account if you want to scale your iteration ceiling.' The architecture choice avoids the iteration-versus-credit conflict at the design level.

The chat stream surfaces a credit_exhausted event. The classifier in route.ts at lines 48 to 68 triggers on Anthropic's structured signals (errorType billing_error or rate_limit, HTTP 402 or 429) and on a regex fallback that matches messages like 'credit balance is too low' or 'out of extra usage.' The PostHog server logs the event (acp-chat-store.ts:971), the UI shows the resetsAt timestamp that the SDK forwards, and the user knows their next turn will work once their Claude plan resets. Critically, the project itself does not die: the sandbox keeps the code, the chat history keeps the turns, and the next attempt picks up where the cap interrupted.

Depends entirely on how much you iterate, and on what you already pay Anthropic. If you iterate a lot and you already have a Claude subscription, the no-markup pass-through is cheaper because there is no platform tax on each turn. If you barely iterate, a small credit pack on a competitor may be cheaper because you never hit the ceiling. The point of the page is not 'mk0r wins on price,' it is 'the incentive structure under the price determines whether the tool encourages or punishes the iteration loop you came for.' Price comparison is downstream of that.

Four files. ANON_TURN_LIMIT and the anonymous gate at src/app/api/chat/route.ts:23 and :193-209. The OAuth-first credential resolution at src/app/api/chat/route.ts:217-228 (loadAndRefreshTokens runs before any shared-key fallback). The credit_exhausted classifier at src/app/api/chat/route.ts:48-68. The user-facing modal copy at src/components/claude-connect-modal.tsx:124-130 ('Use your own Claude subscription to power the AI agent. Your credentials are encrypted and never shared.'). The repository is open source at github.com/m13v/appmaker; you can grep any of these symbols and they appear exactly once each.