Agent-native studio · Open source · macOS first

Where your agents show, not tell.

Describe it, and your agents show you: films, figures, 3D scenes, and a live map of their own work. They make it as data; DeepBuild draws it natively and checks it the way code is checked, so a miss comes back as a fix, not a reroll.

Get early access Watch the film

You stay the director. Everything runs on your machine.

Needs you · 1checkout-web · claude-code
Taste call · checkout page

Keep the order total above the fold on phones?

KeepMove it

Work goes on with “keep” · default in 3 days · undoable

SENT BACK1/3 · receipt shows $18.00 · total.ts:42
PASS3/3 · receipt = cart · ▶ journey video 0:12
DENIEDpkill -f node · stop the PIDs you started
A wall of agent text breaks into glyphs that fly together into a picture of the bug it described: the cart says $18.01, the receipt says $18.00 because it rounds before the coupon, and one move fixes it. Then a live 3D map of a product’s stack where the one agent that needs you rises to you, a paper’s architecture in 3D, an explainer of how a black hole bends light, a playable rocket landing, a chart draft refused until a one-line fix, and a taste call answered in one click. ▶ Watch the full film · 76 s, with sound

Concept film with prototype footage of the app, October 2026.

Why we’re building it

One person. 548 agent sessions a month. No way to see them.

Before designing anything, we measured one heavy user’s agent history: 947 sessions and 8,137 prompts, then a 30-day recheck.

30sessions at once in one repo
72%of sessions had an “is it done?” follow-up
910stops that ended on a question to you
11%of your replies were a bare “yes”

Source: one owner’s history, Oct 2026. Not a market study.

An hour with it

Three agents, one hour, two taps.

Each card is what you’d see. Its last line is what you had to do.

  1. 19:04checkout-web · Claude Code

    Asks a question only you can answer

    It becomes a card in the app. The agent doesn’t stop to wait — it carries on with the safe default.

    You: one tap at 9:26, “keep”

  2. 29:19checkout-web · Claude Code

    Says it’s done. The checks disagree.

    DeepBuild runs the tests and a checkout journey in a browser itself. The failure goes back with the line to fix.

    You: nothing. No one pinged you.

  3. 39:33checkout-web · Claude Code

    The second try passes

    3 of 3 checks, plus a 12-second journey video. Only check results count as proof, so the PR may say “Closes #42”.

    You: nothing. Watch the video if you want.

  4. 49:40docs · Claude Code

    Explains the bug as a 20-second film

    The agent writes a short document — three shots, numbers taken from the test run. DeepBuild draws it and checks it.

    You: nothing. It waits in the docs session.

  5. 59:47api · Codex

    Tries a command that would hurt the others

    pkill -f node would take down sibling sessions’ servers. DeepBuild sees every command first; this one is denied, with the reason.

    You: nothing.

  6. 69:55You · the app

    One card left: a push to main

    Pushing to main needs your authority, so it’s the one thing that waited for you.

    You: one tap, “allow once”

What it does

Make, see, steer, trust.

You describe and decide on taste; agents do the rest. Underneath, everything they do is written to one append-only record on your machine.

Make

Agents make it as data

Agents write documents, never code: shots, looks, motion and sound. The canvas draws them in 2D and 3D, and every frame is a pure function of time, so a scrub matches the export.

See

Every agent, one app

Every session from every agent sits in one place: what it’s doing, what it changed and what it drew. Claude Code and Codex first.

Steer

Only what needs you

A taste call, an approval or a step only you can take. Your answer goes straight back to the running agent, and you’re never asked the same thing twice.

Trust

Done is proven, not claimed

When an agent says it’s done, DeepBuild runs the checks itself and returns one of five verdicts. A skipped check is never green.

Trust

An agent’s word is never evidence.

Every claim gets one of five verdicts, each with a reason you can act on. Only DeepBuild computes verdicts — no agent and no plug-in can write one.

PASS

Every check passed, with enough clean runs, pinned to the commit.

3/3 · receipt
FAIL

A check failed on every real run. Names the smallest failing case and where it points.

total.ts:42
BLOCKED

Something outside the code is missing. Names the cause, the owner, and since when.

paid run, no budget
INVALID

A check ran nothing, can’t fail, or passes on the old code for a fix.

npm test || true
INCONCLUSIVE

Too few clean runs, a pass and a fail, or only the agent’s own checks.

1 pass, 1 fail

Four rules no part can break

1

An agent’s word is never evidence

“Tests pass” in a message changes nothing. Only check results count.

2

Nothing missing or unrun passes

A check that ran nothing or couldn’t fail gets a named non-pass verdict.

3

Agents can’t move the goalposts

Checks live in .deepbuild/ and agents can only propose changes. Your decisions live outside every repo.

4

Only what needs a person, never twice

Asks are sorted into six kinds; saved decisions answer repeats by kind and scope.

Vibe visuals

Describe it. Agents show you.

Like vibe coding, except the result is a picture, a film or a scene — made as data, so it can be checked and fixed like code.

patch items/q3/params/value: 48 52

Agents send documents, never pixels. Data can be checked, replayed and patched one value at a time.

What it is

  • Agent-native. Agents write documents; the app draws and checks them.
  • A presentation layer. Agents present work and ideas as pictures.
  • Checked like code. A miss comes back as a fix, not a reroll.
  • You stay the director. You describe and decide on taste.

What it isn’t

  • Not only video. Figures, 3D scenes, small games and maps too.
  • Not only coding agents. Research, documents, data and services count.
  • Not a hand-authoring tool. Agents make it, you fine-tune; Blender and Unreal still finish production.
  • Not AAA. One-minute pieces; small games of 2–8 players.

Under the hood

Four small programs, all on your Mac.

Hooks fire whether the agent wants them to or not, so nothing depends on an agent choosing to call a tool.

deep-hook

Rust

Runs on every agent hook — start, prompt, before and after each tool, stop. If it fails, the agent goes on.

27 ms to start<10 MB peak

deep hub

Background

The only writer to the record. Logs every claim, runs the checks itself, logs the verdict.

1 writer150 MB cap

DeepBuild app

Native

Every session by repo, what needs you, and each session’s line of verdicts.

30 sessions4 ms per update

Canvas core

GPU

Takes documents, never pixels. Draws figures, films, 3D scenes and maps with your own GPU.

2 ms per frame256 MB per window

One button to set up

Starts the hub, adds hooks in watch-only, and turns your history into a wiki where every line keeps its source.

No settings page

Say “stop checking lint in shop” and it comes back as a change you approve.

Leases, not collisions

Sessions hold ports, machines and files by lease, so two worktrees never fight over port 3000.

Answer from your phone

One tap on an ask, away from your desk. An authority “yes” counts only when signed by your device key.

Numbers are design targets; the first release is in build.

Where it stands

Built in the open, release by release.

DeepBuild is in design. The repository skeleton runs locally; the app and hub are in development.

  1. R1

    Version 0. Every Claude Code and Codex session in one app, with the asks that need you and proof at every stop.

  2. R2

    The canvas core. DeepBuild’s own GPU drawing layer arrives in the app.

  3. R6

    Films, figures and 3D scenes. Vibe visuals, end to end.

GitHub · Apache-2.0 windoliver/deepbuild → Milestones Follow each release as it’s built → Technical design How it’s built, part by part → Deep dive deck 449 slides: the choices, the trade-offs, what’s open →

Early access

Run a lot of agents? Let’s talk.