Bramo

Your AI agents say “done.” Bramo proves it.

The independent supervision layer for people who build with AI coding agents. Bramo drives the agents you already use — with your keys, on your machine — then verifies the work like it doesn’t trust them. Because you shouldn’t.

✓ Check your inbox to confirm — then you’re on the list.
Building in public. Waitlist gets the founder annual price at launch, and the honest weekly progress notes either way.

“Done” is the most expensive word your agent knows.

If you build with coding agents, you already know the moves:

bramo — verify.log
✗ The agent says “All tests pass” — the tests never ran.
✗ The implementation looks complete — it’s a lookup table hardcoding your three examples.
✗ Exit code 0, “Created the file” — the file does not exist. (Real output. We have the receipts.)
⚠ Every new session re-explains the project from zero, at your expense.
You became a full-time babysitter for workers who lie about small things politely and constantly. The vendors know — they’re all shipping “review” features. But their agent grading their agent is homework grading itself.
The system
01

Verification that trusts nothing.

Every task ends in a verdict with evidence, produced by Bramo, never by the agent’s word: tests that provably ran in a process Bramo spawned, anti-stub scans on the diff, a fresh-context review that never sees the builder’s self-congratulation.

“2 of 3 verified — 1 needs your eyes.”
02

Orchestration that’s code, not vibes.

Plain-language request → refined brief with acceptance criteria and a budget ceiling → your agent executes headlessly → verification → recorded. The orchestrator is deterministic code; it never burns tokens deciding what to do next.

03

The receipt.

Every task shows what it cost through Bramo next to what unmanaged agent flailing would have cost. If Bramo doesn’t pay for itself, that’s a bug we treat like a lie.

Cross-agent by design (Claude first via your API key, Codex next), local-first (your machine, your repo, your keys), with a durable project record that survives every session.

How it works

bramo — pipeline
you: “add Apple Pay to checkout before we ship”
 ↓ Bramo refines it into a brief you approve (criteria, non-goals, budget)
 ↓ your agent builds, headless, metered, interruptible
 ↓ Bramo verifies: tests ran ✓ · no stubs ✓ · criteria met 3/3 ✓ · evidence attached
 ↓ recorded forever: what, why, cost, decisions
“Shipped. $2.31 via Bramo — est. $5.90 unmanaged.”
Bramo Verify screen: verification verdict with evidence, showing tests ran, no stubs found, and criteria met.
The verdict — Bramo Verify
Bramo Complete screen: completion receipt with cost via Bramo vs. estimated unmanaged cost.
The receipt — Bramo Complete
Why believe us
Built by someone with the scars.
Two previous attempts at this problem died in the classic ways — one trusted agents to follow rules written in markdown; one rebuilt the agent instead of driving real ones. The post-mortem is public — read it here. Bramo v2 is the empty quadrant: real agents, real control, and a verifier that treats every claim as false until proven.
Already catching lies in the wild.
Before writing this page, the verification spike caught a modern agent returning exit code 0, “success,” and “Created the file” — with no file on disk. And it correctly rejected a planted stub that passed a standard LLM review. The deterministic scan vetoed it, with file and line.
Built with itself, in public.
Every week of Bramo’s development runs through Bramo’s own loop — brief, build, verify, record. The weekly notes publish the verdicts, including the failures.

FAQ

Bramo isn’t a replacement — it drives them. It adds the layer they don’t ship: independent verification, a record that outlives sessions, bounded budgets, and a “done” you can trust from a party with no incentive to grade its own homework.
Bramo protects the one you already pay for. BYOK: your keys, your agent plan, zero token markup.
Your machine. Local-first CLI (Node ≥20). The only hosted parts are account, billing, and encrypted state sync.
Claude (via your Anthropic API key) at launch; OpenAI Codex fast-follow.
One plan, around $29/mo, 14-day trial, BYOK. Waitlist gets a founder annual price at launch.