How to Get Started with DreamCoder — the AI Coding Harness You’ll Actually Want to Use
There is a moment, last week, when I watched a frontier coding agent chew through a piece of enterprise middleware and realized something had shifted. It was fast—maddeningly fast—but I spent more time steering it than I did letting it work. I kept a second terminal open, watching the logs, hunting the moment something went sideways. It sat in a shoe-rack-sized terminal window. Unreadable. A boring sidebar with a chat pane, a chat pane that looked exactly like every other coding harness I have tried. It worked, yes, but it was not built for me. It was built for whoever writes support tickets about it.
That is why a note I read this week caught my eye. On 20 November 2026, Mickey Petersen, the author of Mastering Emacs, formally released DreamCoder—an AI coding harness that rejects that terminal-window orthodoxy. DreamCoder runs in your browser. It is a tiling window manager, an IDE for the agentic age, and it is genuinely different from the chat-pane-of-the-week you are used to. This is a practical primer on what it is, how to try it, and what to expect on the first day.

What DreamCoder is (and what it is not)
DreamCoder is a frontier coding harness. That is a specific thing: it is not a chat app, and it is not another CLI wrapper around Claude Code or Codex. It is a full workspace that sits in your browser, with sub-agents that run in the background, Model Context Protocol (MCP) servers, shell commands, terminal windows, skills, hooks, and a cache hit rate that the author claims sits at 99%. It is built to run for days, not for a single session.
The design language comes from Emacs. DreamCoder uses a tiling window manager: every pane has a minimum, a comfortable, and a maximum size, and a constraint-satisfaction solver keeps your panes snug instead of letting one window swallow the screen. When you ask for a new pane, the layout rebalances; if it cannot wedge the pane in, it hides it in a window bar you can click to resurface. Keyboard-oriented, mouse-optional: Alt+x for global actions (M-x in Emacs), Alt+c for pane-local ones, Alt+m to activate keyboard mode where you type the number of the named pane, and WASD to move between panes.
There are also workspaces—virtual desktops or window configurations you can snapshot, shuffle panes between, and float panes to the top. A minimap in the top-right of each workspace shows every pane, and each one pulses while it is busied, turns amber when it is blocked, and goes green when it is done. You can hover the minimap to see what a pane is doing, or click through to jump there.
The honest framing matters here: DreamCoder is not a magic box. It is a tool that makes the dispatcher role lighter. You still decide what to build, you still review the diffs, and you still catch the bad moves. The author says it himself in his announcement post: the AI does the work, and the human assesses the results. That is the right division of labour, and it is the reason the interface is built the way it is.
How to try it
DreamCoder is a standalone web application with native clients for macOS, Windows, and Linux, plus a Docker image. The download page at dreamcoder.ai is the place to start; the native binaries are the simplest path. If you want to run it inside a container—or you are on a machine where the native install is awkward—the Docker image is a cleaner route than juggling system dependencies.
Once it is running, the setup is genuinely quick. The user interface guides you through choosing your model. You are not locked into a single provider: DreamCoder loads open-weight models such as Qwen, Gemma, Deepseek, and GLM, which run locally. The author has been unusually specific about which local model earns his trust: Qwen 3.8 27B. He calls it fantastically good for its size, and says you can do a genuinely large amount of software development with it if you are a capable hand. Bigger models do more auto-piloting; smaller machines do the job on a budget. If you are on a cash-strapped rig, turn to the self-hosted-model guides and get going that way.
Import what you already use
One of the reasons I want to try this is that it does not start from zero. DreamCoder ships with an importer for Pi, OpenCode, and Codex configs. A couple of clicks and your model settings—plus, presumably, your accumulated preferences—land inside DreamCoder, ready to go. That is the kind of friction reduction that separates a hobby project from a daily driver. If you already run an agentic workflow through one of those tools, you are not starting from an empty pane. For a closer look at building one yourself, how to build an AI agent with Python is a solid primer on the agent side of the harness.)
I would verify the importer against your own config before assuming it picks up every setting. Test it on a throwaway workspace first: confirm your model choice and your key preferences are actually carried over, then commit to the import once you see the output. That is the same habit I use with every new harness: treat the first-run setup as a small experiment, not a promise.
Your first session
Open a chat pane and give it a task. You will immediately notice the difference in how the window behaves: panes for terminals, panes for chats, a minimap that tells you where the work is happening. Resize a pane by dragging the edge; shuffle a pane into a named workspace; float it to the top when you need it on screen. Use the keyboard mode after you have memorized the layout—it is genuinely faster than reaching for the mouse.
Before you let the agent run long, keep a rollback plan. Snapshot your workspace, keep a throwaway git branch, and kill the agent before you let it run for hours unattended the first time. The author’s own advice for the upcoming pentesting suite is also worth following early on: the AI explores, but you assess the impact. Apply that to your first session too.
Run local models, and what to expect
DreamCoder’s support for open-weight models is the part that will decide whether this becomes a permanent tool for you. The author’s take is that the future is local for a lot of development work, and he has the benchmark intuition to back it up: small models like Qwen 3.8 27B (or its cousin, Flash Next) are now capable enough to do real work on a high-end gaming rig or a Mac. If your hardware is older, the self-hosted guides will tell you what is possible. I would not promise you that a 27B model will carry a full-stack enterprise migration on its own—it will not—but it will carry a surprising amount of everyday engineering, and that is the honest benchmark.
For anyone already running local models, the natural next step is to read how to run open-source AI models locally with Ollama; the two fit together in exactly the way DreamCoder assumes. And before you hand any agent a live repo, the vetting guides on AI coding agents are worth a read.
Where this goes next
The announcement ends with a quiet promise about the near future: a pentesting and cybersecurity suite inside DreamCoder, built for humans, where the AI explores and you and your team assess the findings. It is early days, and the author is open to feedback, but it is a sensible direction. The tooling is the last mile of the agentic promise, and DreamCoder is trying to make that last mile a window you want to sit in.
The point of a coding harness is not to make the agent invisible. It is to make you visible enough to catch what the agent does. DreamCoder’s window manager, its workspaces, and its browser form factor all push toward that. Whether it becomes your daily driver will depend on how well it handles your own workflows—so try the import, open one session, and see.
Related reading
- How to Run Open-Source AI Models Locally with Ollama
- How to Build an AI Agent with Python
- How to Vet an AI Coding Agent Before Trusting It With Your Code
- Big Tech Is Here to Let Us Down
Bottom line: DreamCoder is the first coding harness I have used that feels like it was designed for the person at the keyboard rather than the person writing support tickets. Try the import, open one session, and decide for yourself.