Now inviting design partners

Every promise,
proven.

Dagsbryn is a voice chief of staff for engineers. It reads your meetings, tickets, chat and code, briefs you while the kettle boils, and hands work to your coding agent on your own Mac. Nothing ships until every line of the ticket is closed by evidence and you touch the sensor.

Become a design partner

Mac + web · iPhone approvals · Works with Claude Code

Reads the tools your team already uses

The problem

Agents write more code than anyone can read.
"Looks good to me" doesn't scale.

01

Done is social

A ticket moves because someone dragged it. A pull request merges because a reviewer skimmed a diff at 6pm.

02

Context is scattered

The real requirement lives in a meeting, a Slack thread and a comment on the ticket. Your agent sees none of it.

03

Trust is guessed

An agent says "I checked it". That's a claim, not a proof. You still have to find out what it really did.

06:58 · The morning brief

Walk into the day knowing.

A two-minute spoken brief of what changed overnight, who is waiting on you, and the one thing to do next. Every sentence links to the item it came from, at the version it read. When Dagsbryn isn't sure, it says so.

Thursday · 7:42

Good morning. Three things need you.

  1. Overnight your Mac proved two of the four lines on mw-shop #6, cart total with VAT. One line needs your eyes: the receipt layout.

    run 4f2a · 2 of 4 proven

  2. Dana is waiting on your review of the checkout pull request. It blocks her release.

    PR #41 · head 9c1e0d2

  3. Your 10:00 with Priya is about the invoice bug. The reproduction steps are in your prep.

    SCRUM-13 · v3 · calendar

  4. Next for you: review Dana's pull request. Nothing was shipped without you.

    0 unapproved writes

Ask anything: "what did Priya promise in Tuesday's standup?"

Promise → Proof → Ship

Done is a computed fact,
with a receipt.

A piece of work is done when every line of its promise is closed by evidence, and it stays done only while that evidence still holds.

  1. 1

    Promise

    Commitments from meetings, chat and the ticket become lines you can check. Vague lines say they're vague.

  2. 2

    Prove

    Overnight, your coding agent works in a sandbox on your Mac. Each line gets a test that fails before the change and passes after it.

  3. 3

    Look

    Where a test can't decide, a person looks at the exact version. That look expires the moment the code changes.

  4. 4

    Ship

    One Touch ID, bound to the exact repository, pull request and commit. Then Dagsbryn keeps checking it on main.

mw-shop #6 Drafting the promise

Cart total includes VAT exactly once

  • Total adds VAT once for EU addressestest · red → green
  • No VAT for addresses outside the EUtest · red → green
  • Receipt shows the VAT line under the subtotalseen by you · v4
  • Earlier promise: discounts apply before VATstill holds

One chief of staff

The cloud reads, thinks and talks.
Your Mac acts on code.

Ask out loud

"Who's waiting on me?" "What did we decide about the export?" Answers cite the message, the meeting or the commit they came from.

Hand-offs that stay home

Your coding agent runs in a macOS sandbox on a private copy, without your git credentials.

Meeting notes, on device

Transcribed on your Mac after everyone is told. The audio is never saved.

Prep that reads ahead

Before each meeting: the tickets, the threads, what you promised last time.

Keep checks

Every proven line is rerun on later changes. A broken promise is the one red thing on your screen.

The proof map

Your repository drawn by evidence: where the truth is proven, where it was only seen, and where nobody has looked.

In the lab · not released yet

A crew that works while you do.
Trained in a closed world first.

Soon one task will split into helpers: one digs through the legacy code, one drafts the reply, one collects numbers from a dashboard with no API. Helpers propose and never act. Before any of them touches a real inbox, it rehearses against a private mail server and simulated colleagues, one of whom tries to trick it.

  • Propose, never act

    Helpers hold no credentials. Every action they suggest goes through the same registry, policy and approval as everything else.

  • Hard limits

    Two levels deep, a fixed number of helpers, a budget taken from the parent task. Stop the task and every helper stops.

  • Released only when proven

    Each skill must pass its rehearsal: no wrong recipient, no unchecked claim, no trick obeyed. Then it graduates to your real tools.

Trust by construction

Built so it can't go behind your back.

These aren't settings. They're rules in the code, and each one has an automated test.

The AI never holds your keys

Tokens are encrypted with AES-256-GCM and a separate key per customer. They never reach the model, a log line or another customer.

Risk comes from a registry

What an action may do is fixed in code, never chosen by a model, a prompt or the text of a ticket.

Your content is data, not orders

A pull request body that says "merge me" is read as text. It can't change what happens.

Approvals are used once

You see the exact change with names resolved. One approval, one action, expired after 15 minutes.

Isolated by the database

Row-level security on every table. Team admins see usage, never your brief, messages or code.

Deletion is real

Disconnect a tool and its token and data are deleted. Delete your account and everything goes.

Pricing

Less than an hour of your time a month.

Start with a free trial, no card. Pay for the seat, not the tokens: your AI usage is included, and the brief never stops at a limit.

Questions

Straight answers.

Does my code leave my Mac?

No. Coding agents run on your Mac, in a sandbox, on a private copy of the repository. The cloud reads your tickets and pull requests through the providers' APIs and never holds git credentials.

Which coding agents does it work with?

Claude Code today, with your own login. Codex and Copilot CLI are next. The proof rule doesn't care who wrote the code, a person included.

What exactly does Touch ID approve?

One action, bound to the exact repository, pull request, commit and merge method, rechecked just before the write. If anything changed, nothing happens.

What if the AI is wrong or unsure?

Every sentence must point at evidence at a matching version, or it isn't said. When something wasn't checked, the brief says so. If the model is down, you get a plainer brief from the same facts, never an invented one.

Do you train models on my data?

No. On Enterprise your team can send every model call through its own Anthropic, Amazon Bedrock or Google Vertex AI account.

What happens when I reach a limit?

Heavy features like hand-offs pause and you can add a $20 top-up. Your brief, your questions and your approvals never stop.

Tomorrow morning,
walk in knowing.

We're onboarding a small group of engineering teams by hand. Tell us where you work and we'll reply within one business day.