4 agents working

Project canvas for agent swarms

Every project gets a canvas. You talk to one manager, a swarm of agents works the project in parallel on your own machine and subscriptions, and everything they make lands where you can see it, steer it and ship it.

Harness
  • Claude Code
  • Codex
  • Grok Build
  • Muse Code
  • goose
Forge
  • Cursor Origin
  1. Ask the managerOne message. It plans the work and splits it up.
  2. The swarm fans outEach part gets its own agent, sandbox, branch and model, all at once.
  3. The work lands on the canvasImages and clips in the space, builds in Play, progress on the board.
  4. You pick and publishAccept, answer, publish. The swarm keeps going.

01The swarm

One manager. A whole swarm.

You talk to the manager. It splits the work and hands it to subagents that run side by side, each in its own sandbox. It never does the work itself, so it is always free to listen.

  • Delegates, then listens

    Hand-offs return at once. The manager subscribes to pull requests, CI and the clock, and wakes when something it waits on happens.

  • Own clone, own branch

    Every agent gets its own checkout, branch and session in its own sandbox. No worktrees, nothing shared, nothing to clean up.

  • Any harness, any model

    Each agent picks its own: Claude Code, Codex, Grok Build, Muse Code or goose, mixed in one project.

  • Watch the run

    A light per agent while it works. Hover it for the whole run in lanes, with every hand-off, tokens and cost. Steer or stop anyone.

02Spaces

One canvas for people and agents

A space is the project's moodboard: images, notes, links, papers, video and music on a canvas you pan and zoom. Link it to a project, or mention it as #lantern-art, and its agents see it too. They search it, add to it and generate for it.

  • Generate in place

    Right-click the canvas: images, video and music from your own plans. GPT Image on ChatGPT, Grok Imagine on SuperGrok, or a model on your own GPU.

  • Frame by frame

    A real video player. Step frames, split at the exact frame, turn any frame into a reference.

  • Papers, read

    Drop a PDF and its text is extracted and searchable by page. Clip a passage into a snippet with its page.

  • #mention it

    A space in a message reaches the agent at that version: its newest items, notes and images.

  • Agents propose

    What an agent adds lands on an arrivals shelf with its name. Agents can archive. Only people delete.

03Play

Every build gets an address

What agents make opens in Play on its own address: pages, full-stack apps, 3D models, video. No deploy, no sandbox kept awake. Draft is what the agents last wrote. Live is what you published.

  • Apps with a backend

    React and Tailwind in front, a Durable Object with its own SQLite behind. Draft and Live keep separate data.

  • You publish. Agents don't.

    Publish takes a restore point first. Unpublish turns Live off; nothing is lost.

  • Share with a link

    A link shows the running app, never its source, chat or data, and expires in 30 days unless you say otherwise.

  • Views the agents build

    A storyboard, a valuation, a playtest log: agents write the project its own panels, and the manager works them through typed tools.

  • 3D, in place

    GLB, glTF, OBJ, STL, PLY and Gaussian splats open in dr4ft's own viewer.

04Your stack

Your machine. Your plans. Your keys.

dr4ft runs agents on compute you bring, with model subscriptions you already have. The control plane lives on Cloudflare. The work and the logins stay with you.

Bring a machine

$ curl -fsSL \
    https://api.dr4fts.com/machine/install.sh \
  | sudo DR4FT_CODE=K7QX-2M bash
✓ enrolled   studio-box · 32 cores · 128 GB
✓ connected  sandboxes on gVisor (runsc)

Each agent gets its own gVisor sandbox: a user-space kernel between its code and your host. Idle sandboxes sleep, and memory grows with the work.

Use your plans

  • Claude CodeClaude subscriptionsub ✓
  • CodexChatGPT subscriptionsub ✓
  • Grok BuildSuperGroksub ✓
  • Muse CodeMuse subscriptionsub ✓
  • gooseDeepSeek, Kimi, OpenRouter, Gemini or any OpenAI-compatible serverkey

Several logins per harness if you like. Each project uses only the accounts you authorize for it.

Keep the keys

Logins live on your machine, never in images, the database or the cloud. They reach a sandbox only while an agent runs. Code lives on Cursor Origin.

05Proof

Done means checked

A model saying it's done isn't proof. Screenshots and recordings go with every report, your checks run when a turn changed code, and the receipts come from what actually ran.

game-builderreported done4m 12s

The lantern now lights a six-tile radius and dims as fuel runs down. Fuel spawns on three tiles in ten. The HUD shows the bar.

  • npm run checkexit 018s
  • npm testexit 0 · 214 passed41s
  • npx playwright testexit 0 · 6 passed1m 3s
pushed dr4ft/game-builder→a3f9c21 on Origin·PR #41checks passing
playtester's browser You have control
localhost:5173/play
Watching changes nothing. Take control for a login or a captcha.
  • Proof on reports

    What the agent's browser saw goes with its report: screenshots and recordings, kept in the project files.

  • Stop checks

    Your commands, run by dr4ft in the agent's sandbox at the end of its turn. Agents can't change them.

  • Receipts

    Commands with their exit codes, files touched, branches pushed and the head on Origin, read from the turn itself.

06Night shift

It keeps going while you sleep

dr4ft is built to run lightly supervised. Schedules, webhooks and pull-request news wake the manager. Checks that find nothing stay out of your chat. When something needs you, it asks, and your phone buzzes.

  • Schedules and webhooks

    Every night at two, or a POST from any service: each wakes the manager with what it has to say.

  • No wait ends unheard

    Subscriptions carry deadlines: CI that hangs or a review nobody writes still wakes it.

  • Quiet by default

    A check that finds nothing new never reaches your chat. You see what changed.

  • One switch

    Pause the whole project. Nothing runs until someone resumes; deployed apps stay up.

07Primers

Plug in a whole field

A primer is a versioned kit for one domain: skills, specialist roles, tools, a sandbox image built and tested ahead, and views. Attach the ones a project needs. It carries only its own field.

web-games@3attached to Lantern

Browser games in Three.js. The manager plans in the Stage view, one game-builder per part. Playtesters judge builds blind; the art-director runs the ship review.

Roles
game-builder · playtester · art-director · asset-maker
Image
web-games: Three.js, Playwright, the game probe
Views
Stage

And the rest

  • Live together

    Invite people by email. Everyone sees the same threads, live, with who said what.

  • The board

    What was asked, who's on it and what was proven, kept by dr4ft from what happened.

  • Coordination systems

    How agents work together, as versioned files: role cards, skills, tool texts. Pin one, compare two.

  • Rules, memory, skills

    Plain files in the project. What agents write as instructions waits for a person's yes.

  • Bring your client

    Reach your projects from Claude Code or any MCP client with a personal token.

  • Live previews

    An agent's dev server opens in the panel, on its own address, while it works.

  • Autoscaling

    Agents make no resource calls. A sandbox gets what its work takes, within what you give dr4ft.

  • A Mac app

    dr4ft in its own window, with notifications on every device you turn them on.

Give it a project.

dr4ft is by invitation for now. If you have one, your projects are waiting.