●studio · previewgreenfield + brownfield · same agents

01 AI Product Studio

Idea in. Product out.
Watch every step.

Two co-equal entry paths. Greenfield takes plain-language intent through PRD, stack, code, tests, and deploy. Brownfield imports your repo, audits architecture / security / quality, then enhances or renovates per your direction.

Add knowledge base (optional)

Upload documents the agents can use as build-time context (RAG). Works the same in greenfield and brownfield. PDFs, Markdown, and plain-text files work best; the server validates each file and reports any it can't use.

Up to ~20 MB per file (the server is the final word on type + size).

Autom8 is the only fully-wired oversight mode today. Other modes are coming soon.

$0.05 – $50.00 · pre-flight gate (builds refuse to start, never killed mid-flight).

Empty = your plan default.

Auto lets FORMUL8 pick the best-fit stack. Choose one to override.

BYOK — bring your own provider keys (optional)

Use your own LLM key for this build — it lifts the monthly $ cap and is sent only in this request (never stored on a server, never echoed back).

●02 / how it actually works

No magic. Just gates.

Every Studio build flows through eight agents. Validation runs at three tiers; the circuit breaker terminates non-converging retries. The trace is the artifact you ship alongside the code.

/01

Cheap checks first.

T0 is a deterministic checklist that runs in milliseconds — does the file parse? are the imports valid? is the manifest well-formed? A T0 failure aborts before a single T1 sandbox boots.

/02

Sandbox the slow ones.

T1 runs pytest in a sandbox per feature. Failures get one or two more attempts — but only if the scorer trajectory is improving. Same-state retries cost nothing.

/03

Two reviewers, then gate.

T2 sends every passing build past an adversarial AI reviewer with a counter-example prompt. If neither reviewer flags it, the build crosses the gate. If they disagree, a human is paged.

/04

Spend bounded by construction.

The cost meter checks every step against the pre-flight estimate. At 1.5× the estimate the circuit breaker trips — your tokens go where they compound, not where they spin.

●03 / brownfieldyour repo · your keys · your codebase
/03 Brownfield

Point us at your repo.
Get an audit first, code second.

No rewrite tax. The same agents that build greenfield also import an existing codebase, run a three-axis audit, and propose enhancements scoped to what you actually asked for.

  • ▸ Architecture: dependency graph, hotspots, dead modules
  • ▸ Security: secrets, injection surfaces, IAM diffs
  • ▸ Quality: test coverage, type coverage, churn vs coverage
  • ▸ Output: PR-shaped diffs with prompt receipts
audit.reportgithub.com/acme/api-server
  1. arch · ok84/100
  2. sec · warn62/100
  3. qual · fail41/100
· 3 modules with churn > 0.8 and coverage < 0.3· 1 secret in .env.example (committed accidentally)· /search endpoint added · 1 PR · 84 LOC · 12 tests

Watch a build unfold, end to end.

Launch a buildOpen a finished trace