Everything you'd ask,
answered upfront.

It can't. Every run has a hard per-run spend ceilingenforced between phases. Every LLM call is metered. If an agent overshoots its phase budget, the orchestrator aborts the run and you're only charged for what was actually consumed. No surprise bills.

No. The generated app lands on your GitHub, deploys to your Vercel, and writes to your Turso. If you cancel, you keep the repo, the live URL, the database, the hero images, the marketing copy. Moonshift is the orchestrator. The product is yours.

A live deployment URL on your Vercel, a real repo pushed to your GitHub, a 270-character X post and an 800-1,300 character LinkedIn post drafted in your voice, three publication-ready hero images, and a one-click approval gate. You leave with a launch kit, not a sandbox.

Median 6m 42s across logged beta runs. 14 agents run across a 10-phase DAG - planner, db, backend, frontend, tests, contract-validator, deployer, auditor, auditor-security, marketer, image-gen, publisher, plus fixer and audit-fixer retry loops. Phases 4 and 7 fan out in parallel; the rest run sequentially.

Never. The publisher agent writes a PENDING.json and exits. Nothing posts until you approve each post, per platform, from the dashboard. This is a hard product invariant. It can't be bypassed.

Next.js 14 + Drizzle + Turso as the default. Tests in Playwright. 9 starter shapes today: ai-app, api-only, content, dashboard, forms, marketing, saas, scheduling, and a generic base. The planner picks the right shape from your prompt; three are also surfaced as one-click quick-pick tiles in the UI.

A contract-validator diffs the produced code against the planner's spec; a fixer loop retries up to 3x before deploy is even attempted. A separate auditor and auditor-securityscan for quality and vulnerability issues. If it still fails, runs are fully resumable. You don't re-spend Moons on earlier phases.

Yes. Every run bootstraps a persistent project. Each follow-up message becomes a turn routed through a subset of the agent DAG (tweak / feature / refactor / revert). Agents commit to a scratch branch; your main branch is never touched mid-turn. Any turn is revertable in one click.

OpenAI (gpt-5.4-mini and gpt-5.4-nano for reasoning, gpt-image-2 primary plus gpt-image-1.5 fallback for hero images) and Anthropic (claude-sonnet-4-6 for codegen, claude-opus-4-7 for validation). Accounts are configured so your prompts are not used to train third-party models.