Mulberry

Governed AI loop builder · invite-only beta

Describe the job. Get a governed AI worker.

Mulberry is a governed AI loop builder. Describe a job in one sentence and it writes the rules for an AI worker: what it may touch, what it must never do, when it stops, and when it must ask you first. You review the rules, then download a pack you run yourself. Nothing runs here and nothing is stored.

Your email goes to our email provider (Resend) only so we can tell you when access opens. Mulberry stores nothing itself, and you can unsubscribe any time.

  1. 1Describe the job

    One plain-English sentence is enough.

  2. 2Answer safety questions

    One or two, only when the task is risky.

  3. 3Review and tune the rules

    Plain English on the left, editable rules on the right.

  4. 4Export and run

    Download the pack and run it in Claude Code, Codex, or any agent runner.

Generated first

Say what you want. Mulberry drafts the whole , not a blank builder.

Editable second

Tune permissions, , , and . The updates as you edit.

Governed always

A loop cannot be exported until it has a goal, limits, a , and a .

Security is the product

Every Mulberry loop is generated with guardrails: a budget, a stop condition, a permission boundary, a checker, a failure policy, and human approval gates for risky actions. The model only produces structured data. The files in your pack are rendered deterministically, so what you download always matches the rules you reviewed.

Start from a template

View all

Common questions

How to run your pack
Does Mulberry run the AI for me?

No. Mulberry writes the rules and the files. You download the pack and run it in a tool you control, like Claude Code or Codex. Nothing executes here.

Where does my data go?

Your request goes to the model once to draft the contract, then it is discarded. Nothing is stored server-side, and there are no accounts. If you join the waitlist, your email is held by our email provider (Resend) only to notify you about access, and nothing else.

What exactly is in the download?

A zip of plain text files: loop.yaml, the machine-readable contract, plus human-readable policy files for permissions, budget, stop conditions, human gates, failure policy, security notes, and a runbook that tells you what to type first.

What is a checker?

A second, independent agent that must approve the maker's work before the loop can claim success. Medium and high risk loops require one.

Do I need an API key?

No. The builder runs server-side, you never paste a key, and Mulberry never asks for secrets of any kind.

What happens after I run the loop?

Paste the audit-log.jsonl your loop wrote back into the run debrief. Mulberry shows what it did, how each iteration went, and what it cost; add the pack's loop.yaml to check the run against its budget, checker, and permission boundary. Nothing is uploaded; the log is read in your browser.

Around it goes, until the work gets done.