Mastering Emacs author releases DreamCoder, an AI coding harness

Mastering Emacs author releases DreamCoder, an AI coding harness

The author of the Mastering Emacs site has announced the formal release of DreamCoder, an AI coding harness. They say they have been busy with it for the past fifteen months, which explains why the site has been quiet. The post describes DreamCoder as "an AI coding harness that is very different from the ones you're probably familiar with already" and promises both the project announcement and "a potted history of AI and Emacs".

The origin story starts in May 2025. The author installed Claude Code a few days after its first release, put "some coins in the slot machine" and set it to build a complex piece of enterprise middleware they had built several times in their career. The result was far from perfect and needed a lot of human steering, input and correction, but the speed advantage was, in the author's words, undeniable.

What bothered the author was the interface. Claude Code was the first terminal application they had used in a long time, and they found it an odd choice for a coding agent, until everyone else copied it with their own terminal agents. The author's complaint is that the terminal is still the primary modality for most of these tools, and that the GUI versions, which most harnesses now offer, are second-class citizens. They also say the GUIs all look alike: a boring sidebar with chat headings and a chat pane. A serious coder needs more than that, the author writes, and even a casual user will bristle when copying and pasting multi-line text in a terminal.

So in May 2025 the author decided to build a better harness, assuming that with AI it would take no more than 2 or 3 months. It took 15 months, with endless UI and backend iterations, ending in a successful beta test. The author now calls DreamCoder, in their opinion, "the acme of what a coding harness should be".

Before getting to the product itself, the post turns to history. John McCarthy, an AI researcher at MIT, set out in the 1950s to build the Advice Taker, software meant to hold what a basic human should know (he called it commonsense reasoning) in mathematical logic. To do it he created LISP, with M-expressions for functions and S-expressions for symbolic data. Steve Russell, who worked with McCarthy, wrote the first LISP interpreter with a working eval function, and M-expressions were sidelined. McCarthy later conceded that LISP may owe its survival to the fact that its programs are lists. McCarthy never built the Advice Taker in its purest form: the axioms multiplied faster than anyone could write them, and the author compares writing such rules to Prolog.

The thread leads to Emacs. Stallman and Steele created EMACS (Editing MACroS) in the late 1970s at MIT's AI Lab, on top of the inscrutable TECO editor, so they could automate typing tedium. Warren Teitelman, inspired by McCarthy's idea of advising systems, built the ADVISE system in his thesis "PILOT: A Step Toward Man-Computer Symbiosis", a way to intercept a function call and insert code before or after it. MIT hackers later put ADVISE into their own Emacs versions, and it survives in Chalupsky's 1992 Emacs Lisp re-implementation, advice.el, one of Emacs's advice systems today. The author calls it an unlikely lineage from a 1950s symbolic idea to a way of changing functions in GNU Emacs.

The author's conclusion about that history is blunt. SHRDLU, Minsky's frame systems and even M-x doctor all used symbolic "AI", and none brought us closer to the AI its proponents hoped for. Symbolic AI dissipated into the wider body of computer science as tools, languages and concepts. A cast-iron system made of stern rules cannot spawn general intelligence, the author writes, while a ductile system shaped by its master's desires survives because it stays useful. The available text ends mid-argument, before it describes DreamCoder's features.

Key facts

  • The author of Mastering Emacs announces the formal release of DreamCoder, an AI coding harness they say is very different from the ones people already know.
  • The author says it took 15 months, against an original estimate of 2 or 3 months made in May 2025, and ended in a successful beta test.
  • The motivation: after trying Claude Code in May 2025, the author found terminal-centred harnesses limiting and GUI harnesses uniform, a sidebar of chats plus a chat pane.
  • The announcement includes a history running from McCarthy's Advice Taker and LISP through the creation of EMACS to Chalupsky's 1992 advice.el.
  • The available text is cut off before it describes DreamCoder's features, price or availability.

Why it matters

AI coding harnesses are a crowded category, and the author's argument is that most of them share two weaknesses: a terminal-first design and GUI versions that look alike, with a sidebar of chat headings and a chat pane. DreamCoder is pitched as a different take, by someone with long standing in the Emacs community. The 15 months, against a plan of 2 or 3, is also a candid data point on how long building with AI can take.

Who it affects

Developers who use AI coding agents such as Claude Code and are unhappy with the terminal or with chat-sidebar GUIs are the evident audience. Emacs users will recognise the author from the Mastering Emacs site and the Combobulate project, which the post mentions.

How to use it

The available text does not give a price, licence, download link or release date for DreamCoder, and it does not describe the interface or supported models. Anyone interested should read the full announcement at the source link before drawing conclusions about what the tool does.

How solid is it

This is a first-party announcement. The claims about DreamCoder, including that it is "the acme of what a coding harness should be", are the author's opinion. The post mentions a successful beta test but gives no size, participants or results, and no benchmark or comparison against Claude Code or other harnesses. The text available here is truncated before the product description. The history section is the author's own telling, offered as context.

Risks and caveats

Nothing about DreamCoder's capabilities can be verified from the portion available. The author's framing that terminals are primary and GUIs are second-class is a personal view of the market. The 15-month timeline is self-reported. Whether DreamCoder is built on or inside Emacs is not stated in the available text.

“One can even conjecture that LISP owes its survival specifically to the fact that its programs are lists, which everyone, including me, has regarded as a disadvantage.”

— John McCarthy, quoted in the DreamCoder announcement on Mastering Emacs