FAQ
Frequently asked questions
Clear answers about wallet credit, usage, subscriptions, and how Tycoon charges for work.
How does it avoid production incidents from bad migrations?
Three guardrails. First: never combine a destructive change (drop column, drop table) with code that reads the new schema in the same deploy — always ship the additive migration, deploy code that tolerates both schemas, then ship the destructive migration in a separate deploy a week later. Second: every migration gets dry-run against a fresh prod snapshot on staging, with row counts before and after compared automatically. Third: migrations are scheduled for low-traffic windows with explicit rollback SQL pre-written. This protocol has zero production data-loss incidents across the Tycoon and SkillBoss deployments that use it.
Which languages and frameworks does it know?
First class: Node.js (Express, Fastify, Hono, tRPC), Python (FastAPI, Django, Flask), Go (chi, fiber), Ruby (Rails, Sinatra). ORMs: Prisma, Drizzle, TypeORM, SQLAlchemy, Django ORM, ActiveRecord. Databases: PostgreSQL, MySQL, SQLite, MongoDB. Queues: BullMQ, Celery, Sidekiq, Temporal. If your stack is less mainstream (Elixir, Rust, Kotlin), the AI can work in it but iteration is somewhat slower while it learns your codebase conventions.
How does it handle secrets and credentials?
Secrets never appear in code or chat. The AI Backend Engineer reads them from your existing secret store (GCP Secret Manager, AWS Secrets Manager, Doppler, Vault, 1Password) via the runtime environment and references them by name. PRs that accidentally include a literal key are blocked at commit time by a pre-commit hook. When an integration requires a new secret, the AI proposes the name and ships the code; you create the secret in your store with the value. This is the same discipline a careful human engineer would follow, just enforced by default.
Can it own on-call or does it hand off?
It can take first-line on-call for issues it has context on — deployed code, known integrations, established runbooks. What that looks like: PagerDuty page fires, AI Backend Engineer reads logs and traces, forms a hypothesis, proposes a mitigation, pages you only if it exceeds its autonomy boundary (touching prod database directly, reverting a customer-visible feature, triggering a refund). Most founders running Tycoon have the AI handle the first 10 minutes of triage, which catches about 60% of routine issues before they need a human.
What about security and auth?
Standard patterns only. Auth goes through WorkOS, Auth0, Clerk, or your rolled auth that has been reviewed. Passwords get argon2 or bcrypt, never home-grown crypto. Sessions use httpOnly cookies with SameSite=Lax and a reasonable expiry. Input validation at the API boundary with zod, valibot, or pydantic. The AI Backend Engineer refuses to ship code that deserializes untrusted JSON into ORM objects or that concatenates user input into SQL. For anything beyond standard patterns (custom encryption, multi-tenant isolation, regulated data), it flags for human security review rather than improvising.