Turn your dreams into code
One window. Every agent.
Dozens of agents on one screen, hundreds under management — each with its reviews, tasks and history kept alongside. It runs on your own machine; you reach it from any browser, at your desk or on your phone.
macOS 14+ on Apple Silicon · Windows 10 and 11 · Linux x86_64 · Docker
view full size
Keeping an eye on one agent is easy. Fifteen or fifty is not. A terminal leaves you flipping between tabs with no real feedback — squinting at slivers of text, losing track of what each agent is actually doing.
DreamCoder gives every agent, terminal and review its own pane in a tiling workspace that runs in your browser. Drag them around, hide or show them, fork a chat into its own window. Lost that conversation from an hour ago? Open the thread in a pane — done. Want to watch progress? Hit track objectives and a pane keeps score.
Prefer the keyboard? Everything is keyboard and mouse driven. A dedicated keyboard mode moves, swaps and resizes panes and jumps between workspaces without leaving the home row — and the workspace minimap puts a traffic light on every pane: running, needs you, or done.
It never forgets where you were. Open DreamCoder on your phone, tablet or laptop and resume the exact layout you left. And the polish isn’t skin deep: underneath is a state-of-the-art agent engine, so you get a first-class interface and a top-flight coding harness.
Why so many agents? Because one or two no longer cut it. AI has permanently changed how software gets built, and keeping a fleet of agents productive takes more than a wall of terminal tabs.
Here’s what that looks like.
01 · keyboard-first
A lot of the day is spent darting between chats, terminals and reviews, so the whole workspace is keyboard-driven. Tap into keyboard mode and use WASD to move, swap and resize panes, and Q/E to jump between workspaces — hands never leave the home row.
view full size
02 · winch
Press Alt+W and every link, table, button and image in a chat gets a letter. Type the letter to act on it: open a link or copy its URL, copy a whole table as Markdown or TSV, or open an image.
view full size
03 · full fidelity
Test your work as the agent builds it, at the fidelity you’d expect from a browser. Rendered LaTeX, Mermaid diagrams and full Markdown in the chat, with a live terminal running your script right beside it.
view full size
04 · the event graph
Every prompt, tool call, result and objective is a node you can walk. Trace a run from the queued input to the last tool result — and pick up where it actually went, rather than reconstructing it from scrollback.
view full size
05 · objectives
Real software takes more than a hastily written list of to-dos. Ask for something big and the agent writes a proper plan: nested objectives, each with the checks that say when it is done. The plan sits front and centre of your work, and each objective moves from in progress to done, with a timeline of when and a note of what was finished.
view full size
06 · show, not tell
When the agent produces an image — a render, a chart, a screenshot of what it just built — it shows up in the chat. Open any of them in a pane of its own, right beside the conversation, at actual size.
view full size
07 · your own agents
Agent Studio describes an agent in a single sentence you can edit: what it is called, what it may do, which model it runs on and where it is available. Tick whole areas of actions or single ones, and publish it when it’s ready.
view full size
08 · anywhere you are
It is one web app running on your own machine, so your phone does not get a stripped-down companion — it gets the same interface, reflowed. Look in on a long-running agent from the sofa, and send the next instruction.
view full size
For macOS, Windows, Linux and Docker.
A tiling window manager that lives in your browser. Give each agent a pane and see dozens at once — no tabs to lose, no alt-tab, no wondering which terminal was running what. Every reply rendered in full: tables, diagrams, code and equations.
Objectives, collections, review state, and an event graph of every prompt and tool call — all kept, all walkable. Pick a run up where it actually went, instead of re-reading scrollback to work out what your agent decided an hour ago.
Frontier hosted models down to small ones on your own GPU, over any standard endpoint. Native support for GitHub Copilot for Business, or bring your OpenAI Codex subscription. It runs on Windows, macOS and Linux, reached from any browser.
Fantastic cross-platform support
Run it on the Windows laptop your employer issued, the Mac at home, or a Linux box in the cloud. It runs anywhere, and you reach it from anywhere through a browser.
A ZIP for macOS, a Windows setup package, a .deb for Linux, and an OCI image for containers.
$ sudo apt install ./dreamcoder*_amd64.deb Reading package lists... Done Setting up dreamcoder (1.0.11) ... ✓ dreamcoder on your PATH ✓ in your applications menu ✓ updates arrive through apt $ dreamcoder
The same build, the same keybindings, the same screen — whichever machine it's running on.
A laptop on your sofa, your phone on the train, a tablet in bed — they all reach the same instance. Tailscale, Cloudflare Tunnel and plain old port-forwarding are first-class.
runs on
Where this is going
One app for the whole of AI coding — agents, reviews, tasks and terminals in a single place. It isn't all there yet. What's there already lives in one window, and more moves in with every release.
macOS
macOS app ZIP
Windows
Windows setup package
Debian · Ubuntu
amd64
Container
dreamcoder:<release>
Files · Registry
every file · changelog · registry refs
Every published file, the changelog and exact-tag registry refs are on the downloads page
Questions about DreamCoder, advice on setting it up, or help with a problem: email us and we’ll get back to you.
hello@dreamcoder.ai