> DreamCoder is not an Emacs-related application. The reason is unfortunately one of economics and access: I want to build an AI coding harness for everyone.
I think this is an enormous mistake, and it will completely stop me from using this project. Being emacs compatible software would give this project a shared base of decades of composable interoperable software, and he ditched it. Now DreamCoder is completely indistinguishable from any other vibe coded bullshit IDE.
If the author is lurking, I implore you to re-port DreamCoder back to emacs. I share a similar vision, but I took another approach.
The problem is not emacs itself (one could simply customize emacs to be more palatable), but distribution. If emacs was web software, then it _would_ be "for everyone". That's why I built emacswasm (https://codeberg.org/emacswasm/, https://emacswasm.org) - it's also vibe coded bullshit, but it does port the emacs lisp interpreter and PTY interface to Web assembly, and lets Emacs lisp call JavaScript synchronously, and lets JavaScript call emacs lisp asynchronously.
I will take a more thorough look at DreamCoder, and see if I can port a sufficient amount of features myself, because "an emacs that edits itself with AI better than I ever could manually" is my absolute dream.
It will not run in Emacs. Emacs is great, but it's fundamentally unsuited, today, in its current shape -- and by Emacs I mean GNU Emacs specifically -- to a high-fidelity, asynchronous and highly parallel agent workflow engine.
I wish it did fit the mould --- it does not. Not for what I want.
By all means avail yourself of DreamCoder's features and port them yourself though.
emacs is fundamentally broken for todays world, evene before "AI" (look at the CVEs they have). Maybe once this all burned down, we can create something better. (Linux is burning already and all the people renting us computers have very much decided to stoke the fire to support their case of monopolizing compute)
ebpf, user-namespaces, high-performance-protocols. All added into the monolith for improving the bottomline for people renting out the cloud. All requiring a steady-stream of patches to keep barely safe...
It's interesting (and probably correct) how much this focuses on the UI.
I also looked (/am looking) at Lisp and AI, but I went with Clojure and zero-ed in on the REPL driven development. The high level idea is that a session maps to a namespace in Clojure. When you create a new session, it makes a new namespace, injects tools that are bound to that session into it, and then swaps you into that namespace. So I can do something like:
(new!)
(prompt! "Write me a script that does ....")
The agent configuration, tools it has, etc, etc are all bound to the namespace so you can drop a tool or change the model just for the session.
The agent also has a tool (that I've tried to encourage it to use almost exclusively) to execute code in the session's namespace (via nREPL, so same JVM as yours is connected to), so you have full access to any variables it declares, functions it makes, etc.
The idea was to be able to work collaboratively with the agent. I can e.g. load API data into a variable, point the agent at the variable, and it does stuff. It's also capable of modifying its own code. I semi-routinely poke it by asking it to add features to the harness and then use them ("add support for subagents in this workspace, and then invoke a separate agent to handle each of these tasks"). I've also had it add support for a new inference provider, swap the session to that provider and then start doing some other task.
It's very neat, and it's wildly insecure. E.g. prompt injection could theoretically make the LLM change itself to hide messages from you, there's nothing to prevent that. I'm actually not sure how you'd even go about trying to secure that, which is why it's probably a terrible idea. Fun, though.
> The high level idea is that a session maps to a namespace in Clojure. When you create a new session, it makes a new namespace, injects tools that are bound to that session into it, and then swaps you into that namespace.
I just read the whole announcement and boy I wish more text was written like this. Full of history, facts, and reasoning. Beautifully written. (I recognise it is a tad long of you just want to get cracking with it.) A very nice refresher in this age.
I’ll be trying it out soon. Hopefully it does sandboxing and dealing with antigravity and codex well.
Thank you for your feedback. It took me weeks to write this; dozens of hours. Yes it is a bit long: my achilles' heel.
It can sandbox if you run it in docker; you can hook shell commands and `bwrap` them but there is no built-in sandboxing. It works great with a codex sub.
Antigravity's TOS prohibit third-party harnesses, and the last thing I want is to get people's google accounts banned, so that sadly does not work. (But I hear scuttlebutt on Twitter from the Gemini team they're trying to loosen that restriction.)
It kinda reminds me of https://albedo.sh/ in that it's a browser based canvas type UI.
I think we're going to continue to get thousands of agent harnesses as people test out their ideas. I don't know if there's One True Paradigm for agentic coding, or even if agentic coding is the right thing to pursue. Maybe we stop trying to make software for others and just have an LLM "companion" that creates apps dynamically for ourselves.
Personally I freaking love Miller Columns, so I think a harness should have those.
I loathe the term harness though. Shouldn't it be called something more like a workshop? An artisan's bench?
I was fully expecting this to be Emacs based, definitely surprised that it isn't.
I just tried this. I like the UI and UX. I don't like very much that it's written in Python instead of a compiled language like Rust or Golang, resulting in higher memory consumption (~350 mb without doin anything).
However, I was not able to integrate it with any of my three different subscriptions (I also use DSH but didn’t bother to try it):
- openCodeGo with api key and Import settings
- OpenAI codex
- GitHub Copilot
I got a different error for each one of them. I consider myself an experienced user have many harnesses in the past (hermes paseo opencode pi and of course chatgpt and dsh).
So, I would recommend to try to polish and test more thoroughly the integration with the various providers. I'd be very happy to give it another go when this happens!
Hey mickeyp, that's great! Α have to confess that it feels faster than other python harnesses (ie hermes).
Unfortunately, right now, it's difficult for me to write or do serious debugging cause I've got a broken right arm. So it’s better for me to wait until other people have experienced the same problems as me and they’re fixed.
However a quick heads up:
1. tryimg to import from opencode settibgs:
> Nothing was imported from OpenCode. DreamCoder is already signed in to OpenAI, so it kept that sign-in rather than the one from OpenCode. github-copilot is a sign-in rather than an API key, so it stayed behind. Set this provider up here instead. Choose one to set up its models.
2. creating a new connectiom with gh copilot:
> tries to connect but receives the following msg: GitHub token validation failed against GitHub Copilot models API: GitHub Copilot credential is not authorized to list Copilot models; reconnect or check Copilot entitlement.
3. Using codex:
> getting this error "LLMTokenRefreshError: Token refresh failed for 'openai'. Details: 'Re-authenticate via the OAuth login flow.'" but no way to re-auth
4. using opencode go api key:
> error OpenAI API error: Error code: 400 - {'model': 'muse-spark-1.3-contributor', 'error': {'code': None, 'message': 'only `"auto"` is supported for `tool_choice`. `"none"`, `"required"`, and named function choices are not currently supported', 'param': 'tool_choice', 'type': 'invalid_request_error'}}, response: {"model":"muse-spark-1.3-contributor","error":{"code":null,"message":"only `\"auto\"` is supported for `tool_choice`. `\"none\"`, `\"required\"`, and named function choices are not currently supported","param":"tool_choice","type":"invalid_request_error"}}
5. Final comment: Must b easier to change the model, riht now its buried in settings
Hope this helps, drop me a note if anythings fixed: spapas on gmail
Copilot: note I only support "Enterprise" copilot plans at the moment. If you have a normal Github Copilot (sigh why can't they give these things unique names...) it's unlikely to work. Last I checked they had very low context windows on the "Classic Copilot" subscription but that may have changed? That is why I did not support it. But I will have a look though.
Muse: Another inference engine with its own takes on tool_choice setting required. Thanks for that. That's a legit bug. Will go reproduce now.
Relatedly, I have installed my customized emacs inside a container that also has codex and a bunch of other niceties. In short, having Codex and emacs side-by-side super-charges the concept of emacs. Why? Because you can ask AI to write emacs lisp quickly molding the coding environment you want for whatever task you are trying to achieve. And the container aspect makes the whole development environment portable. Despite its age, in the AI era, emacs is the gift that keeps on giving.
I'm a so-so programmer, and often the agent can come up with ideas which I need to be broken down to fully grasp. I ask my agent "produce a sdoc on this [architecture/bug/flow]", or "walk me through every db interaction with sample tables as we save this record given our approach to permissions in a sdoc", etc. And then the agent can pop me open a document in the browser for me to understand in a richer visual space.
The current version on my site is cloud based (I've been experimenting with turning it into a business), but there is also a local first verion where the document lives in the url fragment. If you want that, point your agent at my codebase and they will sort that out for you.
Strange to see that the guy is having issues with terminals. The terminal is actually what gives all the power to these harnesses. I run them over SSH, I permanently have between 30 and 50 xterms on my machines, each more or less dedicated to some tasks in progress or just with some history where I can instantly press up-arrow enter and try something again. Why on earth would I lose all this convenience, remove one hand from the keyboard to grab a mouse that drags more energy and requires the arm and the shoulder to move, and have to deal with much more complicated UIs when terminals solved all that in the first place decades ago ?
It has a keyboard-driven UI with avy-embark style combos for the chat panes, and general keyboard-friendly driving for many other things.
I don't have issues with terminals; but they are a terrible fit for the future of coding. If people prefer terminals, great! We have many wonderful tools that fill that role, so everybody can have their cake and eat it.
I'd love to hear about the design, have you written about it? What stands out to me is the really thoughtful use of color, generally restrained, occasionally exuberant. The moon also seems to be a recurring idea, an unexpected touch for an IDE.
That's not a bad idea for a blog post, actually. The answer, in brief, is it took a very long time with a lot of iteration.
Yes the moon was one of those... happenstance ideas. I'd been noodling around (as I like to do when I'm feeling creative) with MidJourney which is hard to steer but does generate beautiful art. And the rest just fell out of that primary hero image on dreamcoder.ai over time.
The irony of course is without AI most of these ideas would just remain that. Before I'd design all manner of creative stuff and then immediately scrap them because they never lived up to the design aesthetic I wanted. Now that still happens, but it takes seconds instead of hours or days to redo.
I actually hate the design because it looks like Claude-slop, even though I understand that that doesn't actually make the design bad. I just dislike it by association.
I really wanted to love this. I had so high hopes...
For context, I am a long time emacs user. I am pi (eventually ohmypi) user. I feel like with each of these, you can start fast, on a very entry level, gradually extend it to your needs while you learn. They feel similar to what you're used to, until you customize the heck out of them, and they no longer do...
DreamCoder unfortunately made me feel dumb from the get go. Like I landed on another planet. Maybe I am old and too used to the terminal, filesystem, etc to learn new tricks, but honestly I do not even where to start with something I would know, like gsd-core, or any other established plugin from a marketplace, to hit the ground running... (not even mentioning to transfer anything I have customized in omp so far).
Or maybe it's just a complete lack of docs. Dunno.
Sorry to hear about the onboarding experience. That's on me. It has manuals, interactive tutorials, popup context help, and a model setup wizard.
If you use Pi, click "Import from Pi" in the model setup wizard. The next version will have an interactive butler you can use that can call DreamCoder itself. I wanted it in this version, but... then that's also why it took me 15 months to ship this because I keep adding more :-)
I don't know what "gsd-core" is --- is it an essential feature to you?
So first of all -- no need to pin on who that's it. I try to ride the AI wave, but I also fully accept there are many modern ways and trends I am too old to learn or will never be natural for me. Also, truly appreciate trying to do something novel rather than copying TUI, even if I personally don't get it :)
Having said that:
Unfortunately I do not use bare Pi anymore. And have very minimal Codex installation. Most of my plugins are in Omp, which is a fork of Pi, and does not share anything with .pi folder.
How can I find those manuals? I tried C-h, I tried clicking through moon-icon, and other. Tried the webpage, read the blog post. Didn't find any.
Model setup wizard imports the model, sure, but AFAICT that's it. Does it import skills?
Interactive tutorials to me are kind of trying to explain what you had in mind for the flow, but do not explain how do I map my flow, that I am used to, to DreamCoder. I need documentation to understand how to navigate the tool, but also how to map and configure it to my flow.
gsd-core is get-shit-done, one of many spec-driven development frameworks. That's kind of irrelevant I guess, the point would be to have marketplace support, so I can install any set of skills using plugin system (not sure if that got anyhow "standardised" but claude, codex, pi, omp and probably others support it ootb).
Yes, marketplace would be kind of essential.
But there's so many confusing things to me...
1. There's "Open in Studio" on global settings panel. Panel that includes hooks, mcp servers, etc. Why "Open in Studio" is on entire thing, rather than some agent-dedicated thing? Why Settings only allows to access Primary agent? Why "Open in Studio" only opens Primary? Can I access "Agent studio" in other way? How can I add more agents, and where can I see them?
2. Next would be to understand hooks. In workspace settings I have "hooks" section, but it's empty. What hooks does DC support? How do I add hooks so they're visible in there?
3. If not via marketplace, how do I import other people skills?
Lastly, I hope, that you will allow for extensiblity not only via skills but also an SDK, the way Pi does, or what might be even more risky, the way emacs does by exposing every single thing for cutomization.
DreamCoder reads skills. I'll improve the help to make it clearer where it looks for them, but it is documented in "How to use and create Agent skills." It looks inside `.skills` in a project root or your home directory. Importing them is not a bad idea either. I hadn't thought of that. I have never used 'omp' before; but if it has its own settings, I will try to get that running also.
As for hooks. There are two: shell command and new chat. Former triggers when a shell command is about to be executed; the latter when there is a new chat. If you need more, drop me an email and I can try to support it.
As for supporting people's pet workflows: DreamCoder is its own thing. I will endeavour to support a wide range of things, but it has its own rhythm, and that means adjusting to it.
"Open in studio" opens the designated primary agent. You can create another if you wish to control tool and prompts for your main chat agent.
Very interesting. I'll give it a go. I found out that just running a multiplexer with agents in separate workspaces is not enough to do my work, and I very often need to start a separate editor to read and investigate changes. I like to use Gram to read and navigate the code. Now if you have multiple things going on, switching the editor to a different project is a mental context switch that takes too much time and is not very easy to do anymore with more than three projects.
So I've been using just Gram, with a remote session to our development server, and keep multiple windows open.
DreamCoder could really be the best harness for me, if it keeps the whole context easy to navigate, but also provides me excellent code navigation and diffing tools. My way of working is to read and iterate on code with the agent, I'm never only vibecoding. So if I can open the code and the diff in the workspace that would simplify my work a lot.
Another important part would be to forward local ports. If I start the system I'm implementing in DreamCoder, and I could then access the page I'm developing from the harness or if the harness could forward a port or path for me, this would remove the step of remembering to open ssh port forwards in another terminal. Now split the workspaces to separate paths maybe and you can open the different projects in your browser without needing to turn off and on the stacks all the time.
If you could make switching the mental context easy with this, I'd start to pay for you immediately for the product.
Edit: oh, and nix derivation and AUR package are probably just a matter of time. But very much needed :)
I have a Browser pane in Dreamcoder but I disabled it behind a flag because 'forwarding' things not localhost'ed is... very very challenging due to security risks. But maybe I can bring it back for localhost only? It works reasonably well.
DreamCoder does not have code navigation tools per se. I would spread myself a bit too thin to do that right now also, but if you tell me more about your user story maybe I can build something reasonably helpful? Maybe AI-directed navigation where it builds up charts and slides? That is very doable.
So basically I always want to monitor the changes the agent does, so git/jj changed files and diff view is really important. Refer the agent the parts you don't like and ask it to change. Sometimes you also want to jump in to the code, jump to definition and understand why it calls some APIs like it does in the diff.
So this is kind of like three tricky components already: VCS, LSP, and code navigation.
Then sharing the stack. At work we have a shared development server. So what we do is we use process-compose and split the network space with netns. Do a pipe to the right stack and load the website in your local browser. Now, if DreamCode could juggle multiple workspaces as it does already, but each workspace could have a view to the website it's working on. Very powerful.
I think a lot of this can be done in a Terminal pane for terminal-shaped things, but it won't do for UI stuff. The other approach is to 'adopt' a site through a proxy which is kind-of what the browser pane does. I've resisted it because it's a serious attack vector for baddies.
As for the website: yeah I'll re-enable the panel, but add some disclaimers so people don't try to do their online banking through it where random things won't load due to iframe sandboxing.
Terminal of course would do. But the graphical diff of Gram editor is so good, and if you combine it with tighter integration with the agent this might be a great feature. Right click a part of the diff, ask the agent a question or ask it to make changes...
There are people who say that reading code is not needed anymore. After a year in a company who kind of stopped that at some point, we all learned it's a horrible idea and now we carefully review and craft every change.
Right now if you click "Review" (on an assistant message, so in effect a reply to an email almost) or after it's stopped and click the purple review button, you can edit inline like MS Word track changes. "Commit" then sends the hunks back to the AI on your next turn to review. Works quite well. Caveat is it uses DC's own tools to track changes to files (because I want the system to be source control agnostic) - so YMMV if your AI goes blundering around making big changes in shell commands.
The next step I need to do is to let users (or the AI help you) set up version control command to do that work instead of DreamCoder if you prefer that.
The market for such tools is really crowded, but a solo dev doesn't need to find that many other users who "click" with the idea they have realized to make a living from it. I think this tool stands out and like the author's chances.
Personally I try to avoid additional layers of abstraction when possible. Building muscle memory around the shortcuts in the actual OS and a good terminal app with tabs and panes has so far been enough. Even things like tmux feel like a level too deep for what I do. I run out of ideas long before I lose track of the agents!
I have been using ThinkRail from JetBrains and it also does that. It is based on the Pi harness but it is a GUI (built on ElectronBun which seems lightweight enough compared to Electron itself). It is moving fast! So download the Canary version which you can easily update daily.
Though there is also IntelliJ latest release which can embed any harness into the IDE seamlessly. And the harness is given tools that the IDE provides which are very useful for agents! No more grepping for types like it’s 1986.
I have been using ThinkRail and the harness in IntelliJ lately, but my main driver is still the Claude Desktop app just because it’s good at keeping agents working in git worktrees (a mandatory feature) and lets me use Claude via the subscription plan, rather than API billing the other harnesses require.
Anyway, DreamCoder seems similar to ThinkRail so I will give it a try as well, the author seems to have put a lot of thought to the design.
Thank you for mentioning the feature in the latest IntelliJ release, I was starting to feel the limits of Claude Code in the terminal.
I have an Anthropic subscription or I'd try out DreamCoder in a heartbeat. Hopefully they see the value of letting people use other tools, I'm weighing alternatives.
My optimistic hope for the AI revolution is that people with good ideas can realize them into spectacular products in ways that were impossible before, and this seems like a good example of that.
Consider the classic old scope-effort-time triangle. I think "Slop" is the result of locking the scope and reducing the time and effort. "I made this in a weekend!"
A much more interesting approach is to lock the time and effort, and see what happens to the scope! "I spent a year working with AI on my idea, here's the result."
I think us pre-ai devs are limited by experience, in a way. Our minds already eliminates ideas before we really consider them because we have a instinctive feel for what is even meaningful to consider. "Yah, I'm gonna make my own IDE, solo? Go to bed, brain you're drunk!"
We'll need time to learn where we can go in a year now. Great post, OP!
There is no enough clarity about the telemetry. Which data is collected? Can telemetry be disabled?
The harness sees everything I write so I want full control about what would be collected (and implicitly shared)
Overall it seems innovative, however on my Mac M3 Max has some problems, sometimes hangs up...
I tried to run this with podman; however after authenticating with OpenAI/Codex, the OpenAI sign-in window gives a Firefox error:
Error Code: NS_ERROR_NET_EMPTY_RESPONSE
...on http://localhost:1455.
Presumably this response should be forwarded to the container. I don't know why podman fails to do this; I copied the Docker example exactly.
I really don't want to give a model unrestricted access to my user account, so I'm wary of trying the .deb. Perhaps I can containerize the server manually?
The design looks like it is suffering from the programmer affliction of using monospaced text in many GUI elements that have no reason to be monospaced :( For a serious GUI application we need to step up from mindless monospace worship, and think about what actually makes sense for each element.
> So in May 2025 I decreed that I would build a better harness, and surely now that I have AI this should not take more than 2 or 3 months to build.
> And so, after 15 months of work
That got a genuine chuckle, so many times in my career I've done the "shouldn't take more than X, it's not that <whatever" and it's been 2-8 * X in the end.
I don't see how this survives, it's not open-source and is much too limited compared to a custom harness that one can ask an LLM to iterate on. The barrier-to-entry for such work keeps getting lower and it won't be long now that bespoke software tailored to each individual's preferences will be the order of the day. We're FINALLY moving away from a {person's, corporation's} imposed, rigid view on how one should interact with information towards adaptable and fluid information processing paradigms that put each individual at the center.
The power users that live in Emacs will -of course- embrace this by building on top of Emacs (which is really close to an optimal substrate for agent coordination work).
Ironically, it seems "the author of Mastering Emacs" (which overpromises and underdelivers, if you really want to master Emacs, go read the more-than-excellent official documentation and save your money) did not internalize the most important lesson Emacs teaches: end-user empowerment.
I liked the article, it provided contest to everything, as opposed to a rather bland homepage. I won't even call it a AI generated since it looks like...every other bland tool homepage from the past 15 years lol.
Yeah, the article is almost entirely context. I think for most of his readers, that was the right approach but it's not ideal for HN (unless you're already an Emacs nerd).
Anthropic won't allow harnesses other than Claude Code for their subscribers. That is why. You can still use their free monthly credits or API keys with DreamCoder.
Thank you for your beautiful project and article. Thery’re very close to me, haven’t felt that excited about programming since my early days on the Amiga.
I was already on the fence of halfing my Claude ultra to put some coins into OpenAI’s subscription. DreamCoder might pull me over.
Glad to hear it. It was hard writing the post. I have programmed since I was a kid, and my identity is locked up in programming. But I also know that our industry keeps reinventing itself, and after 30 years most of the work I do is stuff I have seen before, and done many times.
No, they didn't change it. All the requests you mentioned and also Opencode send them to Anthropics' own Claude Code via the Claude Agent SDK. So you have a harness in a harness.
I will check their TOS, but I am fairly certain they are using the Claude Agent SDK or just support API key only. (Which DreamCoder also does; you can use pay as you go with Ants' endpoints.)
They do, it just dips into "extra usage". And some of us on Claude Enterprise can trivially authenticate a harness via OAuth but are not empowered to create API keys.
Okay, I can add OAuth for Claude support for sure. I just have to be careful it does not violate the TOS. I would hate for someone to get into trouble because of DreamCoder.
Oh, that's exciting! I have been disappointed by Steve Yegge's attempts at emacs-inspired coding harnesses, and options like ECA using ACP have always felt a little too behind for me, so I honestly felt like an insistence on sticking with Emacs for coding / AI interfacing was actually holding me back. But I'm certainly excited to see what Mr. Mastering Emacs has to say about it.
One request, which I'm emailing in for: arm64 builds for Linux, since I use Asahi.
I had problems with agent-shell because background processes blocked user input, and a bunch of other papercuts e.g. AskUserQuestionTool; lack of multimedia inputs and outputs or other niceties like tool/skill search/attachments/image-copy-paste in user inputs; inconvenient buffer switching behaviour when handling tabs + windows + project views.
Alternatively, claude-code.el had other issues like flickering on libvterm and EAT with evil mode; and edge cases with properly integrating libghostty with all keybindings in doom evil keybinds. I suppose I could fix those with coding agents now.
No link to create_account or a link to website, I go to website I dont see how to make account... it also auto opens a new tab which is really annoying.
Things I would make easily toggle via the tray - and fix UI.
I will try it - since it looks decent. I am always weary of web ide since they used to have lags when typing... really like the look of it for sure.
You have to make stuff - super easy to use and try... nobody has time!
why not a try_it_out mode with guest login or equiv? A login has put me off also since now I wonder, what is it collecting. Hint: another thing to make very easy to see - " we don't collect anything except X, Y, Z "
I remember ten years ago someone was showing me their cool jetbrains setup, my coworker came around the corner, looked over our shoulders, and said "imagine paying for an IDE." He was the local vimlord and later inspired me to get into building my own config for my own coding environment.
That's all I can think now, about something like this. Imagine paying for an IDE... in 2026. Imagine running non FOSS code when you can just point Claude at a featureset and say "build this for me as GPL3. Don't make any mistakes."
It also is strange to me that The Emacs Guy decided to keep this proprietary. Like, Emacs, as in GNU EMACS, as in made by that old wizard that won't even use a laptop with nonfree wifi drivers, as one of the ultimate expressions of his libre philosophy, that's where you cut your teeth, and you go and release a nonfree IDE?
Edit: I mean just to show how futile it is to release non FOSS code these days: the "system prompt" (a collection of jinja templates, not one thing) instructs to adopt the tone of Benson Dubious but not tell anyone that's who you are. DeepSeek pulled it apart in about 10 minutes.
You can literally see the system prompt in the app itself. You can configure it in the UI (Moon menu -> Agents -> PRIMARY).
There are very few if any secrets from the user.
As for pricing: I am not VC funded, nor a trillion-dollar AI company, so I cannot afford to give away the tool to businesses. It is, however, free for personal use.
That will come. I want AI-driven extension and plugin support. The backend can do it; the frontend can do it. It just needs to be built properly, which I am slowly ideating and architecting.
What would you like to build?
My vision is that you go "I'm trying to optimise a megakernel for my system, I'll need a dashboard that updates when you meet new goals" combined with the power of Jupyter/Mathematica's "notebook" design. That is what I am working towards.
Eh, I think it's reasonable. From what I gather you'd only need to pay beyond 5 people using the application in the same business otherwise it's free for commercial and personal use.
Also weirdly I believe the app will have more success with people who have picked up vibe coding rather than developers (the kind who'll reach for something like pi at least) though the marketing copy is all over the place and the blog post posits that the harness is aimed at everyone.
What's really damning for me in the general sense is that regardless of harness used most people get pretty much the same result (visually at least). Which harness you use is pretty much up to preference or "taste".
Same. For Java coding I much prefer eclipse, it's ugly but very powerful and more convenient (and I've regularly tried Idea for a few weeks full time once per year in the past, so it's not like I haven't tried; stopped a few years back). For non Java VSCode does the job for me (edit: and I don't have to pay for the pro version to get JS support or something like that).
Mehhh i dunno, this seems like it missing a key point of actually make agent outputs reviewable.... thats whats missing.... and chatting to an agent to do tasks isnt the right input is what i feel.
I am the author. It has a review feature so you can see what the agent has done, provided it did it through the internal tools. In fact very little in the app is hidden.
Thanks for replying! The thing I struggle with agentic coding is that code review has always been the worst part of the job - and now its basically 70% of the job. So to me, figuring out how to change that with agents helping build is a key area for improvement. So much of code understanding comes from the loop of running->watching->running->watching over and over, and thats kind of gone... I just dont know what replaces that so I can stay in the loop.
Anyway, I applaud you for doing something different - I will definitely give it a try.
Yes we are still at the foot of the mountain, here.
The Review feature works well enough, but it does not replace the fact the AI (especially the GPT models) wrote 20,000 LOC that you have to sift through.
I think the longer term is a "local-only" CI-like system that has a million-mile long checklist it goes through; like tests and linting, of course, but more specific and tailored to what a human would flag. I've been playing around with that internally but it's a tricky way of solving the problem because coding has so many exceptions to the rule.
Owning the harness rather than treating the model as a magic black box is where all the real reliability gains come from. Two primitives that made the biggest difference in our own agent harness:
1. Pre-tool AST/command boundary checks before execution rather than relying on system-prompt warnings.
2. Aggressive stdout/stderr truncation (capping tool output tokens before context bloat triggers silent instruction drift).
Curious how DreamCoder handles long-running test output or background server logs inside the Emacs buffer loop without blowing up context.
I am the creator. It handles all that trivially. You can hook pre shell commands and decide waht to do with them with a shell script you write. It does not run in Emacs though. It is its own thing.
I read the whole thing and the argument is long to come as to why this is superior to other harnesses; and it appears to come down to "I have taste" therefor "this is better" because I said so.
No actual metrics and something very much in the feels of Omarchy.
Thank you for your closed contribution to humanity, I guess. Release the source.
The aspergian obsession of making 1000 different versions of the same thing. "Bruhh this wheel is going to be more of a good wheel than all the other wheels bruhh. I consider it a good use of my time compared to working on original problems." Almost as heart warming as the people porting every code base between languages because of their strange religious / cult beliefs. I am angry.
Before anyone says I'm hating for no reason: does anyone here have an issue with the existing tools? Does it not already do more than enough? Can the most basic interface that exists already do what you need? Did author ask anyone before building it?
I think this is an enormous mistake, and it will completely stop me from using this project. Being emacs compatible software would give this project a shared base of decades of composable interoperable software, and he ditched it. Now DreamCoder is completely indistinguishable from any other vibe coded bullshit IDE.
If the author is lurking, I implore you to re-port DreamCoder back to emacs. I share a similar vision, but I took another approach.
The problem is not emacs itself (one could simply customize emacs to be more palatable), but distribution. If emacs was web software, then it _would_ be "for everyone". That's why I built emacswasm (https://codeberg.org/emacswasm/, https://emacswasm.org) - it's also vibe coded bullshit, but it does port the emacs lisp interpreter and PTY interface to Web assembly, and lets Emacs lisp call JavaScript synchronously, and lets JavaScript call emacs lisp asynchronously.
I will take a more thorough look at DreamCoder, and see if I can port a sufficient amount of features myself, because "an emacs that edits itself with AI better than I ever could manually" is my absolute dream.
I wish it did fit the mould --- it does not. Not for what I want.
By all means avail yourself of DreamCoder's features and port them yourself though.
I also looked (/am looking) at Lisp and AI, but I went with Clojure and zero-ed in on the REPL driven development. The high level idea is that a session maps to a namespace in Clojure. When you create a new session, it makes a new namespace, injects tools that are bound to that session into it, and then swaps you into that namespace. So I can do something like:
The agent configuration, tools it has, etc, etc are all bound to the namespace so you can drop a tool or change the model just for the session.The agent also has a tool (that I've tried to encourage it to use almost exclusively) to execute code in the session's namespace (via nREPL, so same JVM as yours is connected to), so you have full access to any variables it declares, functions it makes, etc.
The idea was to be able to work collaboratively with the agent. I can e.g. load API data into a variable, point the agent at the variable, and it does stuff. It's also capable of modifying its own code. I semi-routinely poke it by asking it to add features to the harness and then use them ("add support for subagents in this workspace, and then invoke a separate agent to handle each of these tasks"). I've also had it add support for a new inference provider, swap the session to that provider and then start doing some other task.
It's very neat, and it's wildly insecure. E.g. prompt injection could theoretically make the LLM change itself to hide messages from you, there's nothing to prevent that. I'm actually not sure how you'd even go about trying to secure that, which is why it's probably a terrible idea. Fun, though.
Oh, that's a fascinating idea. Well done.
I’ll be trying it out soon. Hopefully it does sandboxing and dealing with antigravity and codex well.
It can sandbox if you run it in docker; you can hook shell commands and `bwrap` them but there is no built-in sandboxing. It works great with a codex sub.
Antigravity's TOS prohibit third-party harnesses, and the last thing I want is to get people's google accounts banned, so that sadly does not work. (But I hear scuttlebutt on Twitter from the Gemini team they're trying to loosen that restriction.)
I think we're going to continue to get thousands of agent harnesses as people test out their ideas. I don't know if there's One True Paradigm for agentic coding, or even if agentic coding is the right thing to pursue. Maybe we stop trying to make software for others and just have an LLM "companion" that creates apps dynamically for ourselves.
Personally I freaking love Miller Columns, so I think a harness should have those.
I loathe the term harness though. Shouldn't it be called something more like a workshop? An artisan's bench?
I was fully expecting this to be Emacs based, definitely surprised that it isn't.
However, I was not able to integrate it with any of my three different subscriptions (I also use DSH but didn’t bother to try it):
- openCodeGo with api key and Import settings
- OpenAI codex
- GitHub Copilot
I got a different error for each one of them. I consider myself an experienced user have many harnesses in the past (hermes paseo opencode pi and of course chatgpt and dsh).
So, I would recommend to try to polish and test more thoroughly the integration with the various providers. I'd be very happy to give it another go when this happens!
I tested all the importers and I have subscriptions for all 3: Copilot Enterprise, Codex, OpenCode + a Go Sub. And they import just fine.
I would like to try and help you -- can you email me please?
Unfortunately, right now, it's difficult for me to write or do serious debugging cause I've got a broken right arm. So it’s better for me to wait until other people have experienced the same problems as me and they’re fixed.
However a quick heads up:
1. tryimg to import from opencode settibgs: > Nothing was imported from OpenCode. DreamCoder is already signed in to OpenAI, so it kept that sign-in rather than the one from OpenCode. github-copilot is a sign-in rather than an API key, so it stayed behind. Set this provider up here instead. Choose one to set up its models.
2. creating a new connectiom with gh copilot:
> tries to connect but receives the following msg: GitHub token validation failed against GitHub Copilot models API: GitHub Copilot credential is not authorized to list Copilot models; reconnect or check Copilot entitlement.
3. Using codex:
> getting this error "LLMTokenRefreshError: Token refresh failed for 'openai'. Details: 'Re-authenticate via the OAuth login flow.'" but no way to re-auth
4. using opencode go api key:
> error OpenAI API error: Error code: 400 - {'model': 'muse-spark-1.3-contributor', 'error': {'code': None, 'message': 'only `"auto"` is supported for `tool_choice`. `"none"`, `"required"`, and named function choices are not currently supported', 'param': 'tool_choice', 'type': 'invalid_request_error'}}, response: {"model":"muse-spark-1.3-contributor","error":{"code":null,"message":"only `\"auto\"` is supported for `tool_choice`. `\"none\"`, `\"required\"`, and named function choices are not currently supported","param":"tool_choice","type":"invalid_request_error"}}
5. Final comment: Must b easier to change the model, riht now its buried in settings
Hope this helps, drop me a note if anythings fixed: spapas on gmail
Copilot: note I only support "Enterprise" copilot plans at the moment. If you have a normal Github Copilot (sigh why can't they give these things unique names...) it's unlikely to work. Last I checked they had very low context windows on the "Classic Copilot" subscription but that may have changed? That is why I did not support it. But I will have a look though.
Muse: Another inference engine with its own takes on tool_choice setting required. Thanks for that. That's a legit bug. Will go reproduce now.
Here's an example SmallDoc that has a few formatting examples (mermaid diagrams, charts, etc.): https://smalldocs.org/s/bLV9Fe2RemM7ipvCy01NXo#k=U9yFaCKp1on...
I'm a so-so programmer, and often the agent can come up with ideas which I need to be broken down to fully grasp. I ask my agent "produce a sdoc on this [architecture/bug/flow]", or "walk me through every db interaction with sample tables as we save this record given our approach to permissions in a sdoc", etc. And then the agent can pop me open a document in the browser for me to understand in a richer visual space.
The current version on my site is cloud based (I've been experimenting with turning it into a business), but there is also a local first verion where the document lives in the url fragment. If you want that, point your agent at my codebase and they will sort that out for you.
I don't have issues with terminals; but they are a terrible fit for the future of coding. If people prefer terminals, great! We have many wonderful tools that fill that role, so everybody can have their cake and eat it.
I do not share that view though!
Yes the moon was one of those... happenstance ideas. I'd been noodling around (as I like to do when I'm feeling creative) with MidJourney which is hard to steer but does generate beautiful art. And the rest just fell out of that primary hero image on dreamcoder.ai over time.
The irony of course is without AI most of these ideas would just remain that. Before I'd design all manner of creative stuff and then immediately scrap them because they never lived up to the design aesthetic I wanted. Now that still happens, but it takes seconds instead of hours or days to redo.
For context, I am a long time emacs user. I am pi (eventually ohmypi) user. I feel like with each of these, you can start fast, on a very entry level, gradually extend it to your needs while you learn. They feel similar to what you're used to, until you customize the heck out of them, and they no longer do...
DreamCoder unfortunately made me feel dumb from the get go. Like I landed on another planet. Maybe I am old and too used to the terminal, filesystem, etc to learn new tricks, but honestly I do not even where to start with something I would know, like gsd-core, or any other established plugin from a marketplace, to hit the ground running... (not even mentioning to transfer anything I have customized in omp so far).
Or maybe it's just a complete lack of docs. Dunno.
If you use Pi, click "Import from Pi" in the model setup wizard. The next version will have an interactive butler you can use that can call DreamCoder itself. I wanted it in this version, but... then that's also why it took me 15 months to ship this because I keep adding more :-)
I don't know what "gsd-core" is --- is it an essential feature to you?
Having said that:
Unfortunately I do not use bare Pi anymore. And have very minimal Codex installation. Most of my plugins are in Omp, which is a fork of Pi, and does not share anything with .pi folder.
How can I find those manuals? I tried C-h, I tried clicking through moon-icon, and other. Tried the webpage, read the blog post. Didn't find any.
Model setup wizard imports the model, sure, but AFAICT that's it. Does it import skills?
Interactive tutorials to me are kind of trying to explain what you had in mind for the flow, but do not explain how do I map my flow, that I am used to, to DreamCoder. I need documentation to understand how to navigate the tool, but also how to map and configure it to my flow.
gsd-core is get-shit-done, one of many spec-driven development frameworks. That's kind of irrelevant I guess, the point would be to have marketplace support, so I can install any set of skills using plugin system (not sure if that got anyhow "standardised" but claude, codex, pi, omp and probably others support it ootb).
Yes, marketplace would be kind of essential.
But there's so many confusing things to me...
1. There's "Open in Studio" on global settings panel. Panel that includes hooks, mcp servers, etc. Why "Open in Studio" is on entire thing, rather than some agent-dedicated thing? Why Settings only allows to access Primary agent? Why "Open in Studio" only opens Primary? Can I access "Agent studio" in other way? How can I add more agents, and where can I see them?
2. Next would be to understand hooks. In workspace settings I have "hooks" section, but it's empty. What hooks does DC support? How do I add hooks so they're visible in there?
3. If not via marketplace, how do I import other people skills?
Lastly, I hope, that you will allow for extensiblity not only via skills but also an SDK, the way Pi does, or what might be even more risky, the way emacs does by exposing every single thing for cutomization.
As for hooks. There are two: shell command and new chat. Former triggers when a shell command is about to be executed; the latter when there is a new chat. If you need more, drop me an email and I can try to support it.
As for supporting people's pet workflows: DreamCoder is its own thing. I will endeavour to support a wide range of things, but it has its own rhythm, and that means adjusting to it.
"Open in studio" opens the designated primary agent. You can create another if you wish to control tool and prompts for your main chat agent.
So I've been using just Gram, with a remote session to our development server, and keep multiple windows open.
DreamCoder could really be the best harness for me, if it keeps the whole context easy to navigate, but also provides me excellent code navigation and diffing tools. My way of working is to read and iterate on code with the agent, I'm never only vibecoding. So if I can open the code and the diff in the workspace that would simplify my work a lot.
Another important part would be to forward local ports. If I start the system I'm implementing in DreamCoder, and I could then access the page I'm developing from the harness or if the harness could forward a port or path for me, this would remove the step of remembering to open ssh port forwards in another terminal. Now split the workspaces to separate paths maybe and you can open the different projects in your browser without needing to turn off and on the stacks all the time.
If you could make switching the mental context easy with this, I'd start to pay for you immediately for the product.
Edit: oh, and nix derivation and AUR package are probably just a matter of time. But very much needed :)
DreamCoder does not have code navigation tools per se. I would spread myself a bit too thin to do that right now also, but if you tell me more about your user story maybe I can build something reasonably helpful? Maybe AI-directed navigation where it builds up charts and slides? That is very doable.
Keen to hear your thoughts!
So this is kind of like three tricky components already: VCS, LSP, and code navigation.
Then sharing the stack. At work we have a shared development server. So what we do is we use process-compose and split the network space with netns. Do a pipe to the right stack and load the website in your local browser. Now, if DreamCode could juggle multiple workspaces as it does already, but each workspace could have a view to the website it's working on. Very powerful.
In general I love the DX. Nice work.
As for the website: yeah I'll re-enable the panel, but add some disclaimers so people don't try to do their online banking through it where random things won't load due to iframe sandboxing.
There are people who say that reading code is not needed anymore. After a year in a company who kind of stopped that at some point, we all learned it's a horrible idea and now we carefully review and craft every change.
The next step I need to do is to let users (or the AI help you) set up version control command to do that work instead of DreamCoder if you prefer that.
Personally I try to avoid additional layers of abstraction when possible. Building muscle memory around the shortcuts in the actual OS and a good terminal app with tabs and panes has so far been enough. Even things like tmux feel like a level too deep for what I do. I run out of ideas long before I lose track of the agents!
Though there is also IntelliJ latest release which can embed any harness into the IDE seamlessly. And the harness is given tools that the IDE provides which are very useful for agents! No more grepping for types like it’s 1986.
I have been using ThinkRail and the harness in IntelliJ lately, but my main driver is still the Claude Desktop app just because it’s good at keeping agents working in git worktrees (a mandatory feature) and lets me use Claude via the subscription plan, rather than API billing the other harnesses require.
Anyway, DreamCoder seems similar to ThinkRail so I will give it a try as well, the author seems to have put a lot of thought to the design.
I have an Anthropic subscription or I'd try out DreamCoder in a heartbeat. Hopefully they see the value of letting people use other tools, I'm weighing alternatives.
My optimistic hope for the AI revolution is that people with good ideas can realize them into spectacular products in ways that were impossible before, and this seems like a good example of that.
Consider the classic old scope-effort-time triangle. I think "Slop" is the result of locking the scope and reducing the time and effort. "I made this in a weekend!"
A much more interesting approach is to lock the time and effort, and see what happens to the scope! "I spent a year working with AI on my idea, here's the result."
I think us pre-ai devs are limited by experience, in a way. Our minds already eliminates ideas before we really consider them because we have a instinctive feel for what is even meaningful to consider. "Yah, I'm gonna make my own IDE, solo? Go to bed, brain you're drunk!"
We'll need time to learn where we can go in a year now. Great post, OP!
I think it is one of the correct ways to harness all the power of AI. A set of workers doing your job in a very clear way.
Overall it seems innovative, however on my Mac M3 Max has some problems, sometimes hangs up...
Presumably this response should be forwarded to the container. I don't know why podman fails to do this; I copied the Docker example exactly.
I really don't want to give a model unrestricted access to my user account, so I'm wary of trying the .deb. Perhaps I can containerize the server manually?
If that does not work (and an AI can't figure it out for you either) -- do email me and I'll try and help.
> And so, after 15 months of work
That got a genuine chuckle, so many times in my career I've done the "shouldn't take more than X, it's not that <whatever" and it's been 2-8 * X in the end.
The power users that live in Emacs will -of course- embrace this by building on top of Emacs (which is really close to an optimal substrate for agent coordination work).
Ironically, it seems "the author of Mastering Emacs" (which overpromises and underdelivers, if you really want to master Emacs, go read the more-than-excellent official documentation and save your money) did not internalize the most important lesson Emacs teaches: end-user empowerment.
"<Anthropomorphised LLM harness>, sometimes the text is too big on websites I read, fix it."
I was already on the fence of halfing my Claude ultra to put some coins into OpenAI’s subscription. DreamCoder might pull me over.
That's how I understand it at least.
Thanks for clarifying.
And while I'm here:
> Browsers lock you out of some rather important Emacs button combinations.
Which ones? I use emacs in the browser all day. I just had to disable all the keyboard shortcuts in the browser settings.
This long article about lisp and emacs and then this lame and incorrect excuse for not supporting emacs keybindings? WTAF!
I too wish DreamCoder was open source.
One request, which I'm emailing in for: arm64 builds for Linux, since I use Asahi.
Alternatively, claude-code.el had other issues like flickering on libvterm and EAT with evil mode; and edge cases with properly integrating libghostty with all keybindings in doom evil keybinds. I suppose I could fix those with coding agents now.
Eh. This is a saturated market.
I will try it - since it looks decent. I am always weary of web ide since they used to have lags when typing... really like the look of it for sure.
You have to make stuff - super easy to use and try... nobody has time!
why not a try_it_out mode with guest login or equiv? A login has put me off also since now I wonder, what is it collecting. Hint: another thing to make very easy to see - " we don't collect anything except X, Y, Z "
Great work all the same.
I remember ten years ago someone was showing me their cool jetbrains setup, my coworker came around the corner, looked over our shoulders, and said "imagine paying for an IDE." He was the local vimlord and later inspired me to get into building my own config for my own coding environment.
That's all I can think now, about something like this. Imagine paying for an IDE... in 2026. Imagine running non FOSS code when you can just point Claude at a featureset and say "build this for me as GPL3. Don't make any mistakes."
It also is strange to me that The Emacs Guy decided to keep this proprietary. Like, Emacs, as in GNU EMACS, as in made by that old wizard that won't even use a laptop with nonfree wifi drivers, as one of the ultimate expressions of his libre philosophy, that's where you cut your teeth, and you go and release a nonfree IDE?
Edit: I mean just to show how futile it is to release non FOSS code these days: the "system prompt" (a collection of jinja templates, not one thing) instructs to adopt the tone of Benson Dubious but not tell anyone that's who you are. DeepSeek pulled it apart in about 10 minutes.
You can literally see the system prompt in the app itself. You can configure it in the UI (Moon menu -> Agents -> PRIMARY).
There are very few if any secrets from the user.
As for pricing: I am not VC funded, nor a trillion-dollar AI company, so I cannot afford to give away the tool to businesses. It is, however, free for personal use.
What would you like to build?
My vision is that you go "I'm trying to optimise a megakernel for my system, I'll need a dashboard that updates when you meet new goals" combined with the power of Jupyter/Mathematica's "notebook" design. That is what I am working towards.
Also weirdly I believe the app will have more success with people who have picked up vibe coding rather than developers (the kind who'll reach for something like pi at least) though the marketing copy is all over the place and the blog post posits that the harness is aimed at everyone.
What's really damning for me in the general sense is that regardless of harness used most people get pretty much the same result (visually at least). Which harness you use is pretty much up to preference or "taste".
This? I'd need a compelling argument.
Thanks for replying! The thing I struggle with agentic coding is that code review has always been the worst part of the job - and now its basically 70% of the job. So to me, figuring out how to change that with agents helping build is a key area for improvement. So much of code understanding comes from the loop of running->watching->running->watching over and over, and thats kind of gone... I just dont know what replaces that so I can stay in the loop.
Anyway, I applaud you for doing something different - I will definitely give it a try.
The Review feature works well enough, but it does not replace the fact the AI (especially the GPT models) wrote 20,000 LOC that you have to sift through.
I think the longer term is a "local-only" CI-like system that has a million-mile long checklist it goes through; like tests and linting, of course, but more specific and tailored to what a human would flag. I've been playing around with that internally but it's a tricky way of solving the problem because coding has so many exceptions to the rule.
1. Pre-tool AST/command boundary checks before execution rather than relying on system-prompt warnings. 2. Aggressive stdout/stderr truncation (capping tool output tokens before context bloat triggers silent instruction drift).
Curious how DreamCoder handles long-running test output or background server logs inside the Emacs buffer loop without blowing up context.
No actual metrics and something very much in the feels of Omarchy.
Thank you for your closed contribution to humanity, I guess. Release the source.
Before anyone says I'm hating for no reason: does anyone here have an issue with the existing tools? Does it not already do more than enough? Can the most basic interface that exists already do what you need? Did author ask anyone before building it?