Skip to content

DreamCoder brings a tiling window manager to AI coding agents

The Mastering Emacs author spent 15 months building a browser-based harness for running several coding agents side by side, with workspaces, sub-agents and MCP support.

Source: Mastering Emacs

Mickey Petersen, who writes the Mastering Emacs blog, released DreamCoder on 20 November: a browser-based coding harness built around a tiling window manager for running AI coding agents. It runs on macOS, Windows, Linux and Docker, and connects to frontier hosted models or to local models on your own GPU over a standard endpoint, with the product site listing native support for GitHub Copilot for Business or a bring-your-own OpenAI Codex subscription.

The interface borrows ideas from the Emacs world, Avy-style window selection, Embark-like actions, a Helm-style search, without DreamCoder itself being an Emacs extension: Petersen says browser security restrictions rule out full Emacs keybindings. Workspaces act as virtual desktops with a minimap showing task progress across concurrent agents, and the harness supports backgroundable sub-agents and MCP servers directly. Among local models, Petersen singles out Qwen 3.8 27B as “fantastically good for its size.” He started the project in May 2025 expecting it to take two or three months; it took fifteen.

Why it matters: coding agents increasingly need somewhere to run several tasks at once without losing track of what each one is doing, and most editors still treat an agent as a single chat panel bolted onto the side. DreamCoder’s bet is that the agents themselves are commodity now, and the orchestration layer watching several run at once is the actual product.

The caveat: no pricing or licensing is published on the site yet, and the pentesting suite the author previews is still unfinished.

Share

Related reading

More from The Wire

All briefs