Working with LLMs in Emacs is pure joy. LLMs are all about text and Emacs is the absolute king here. Nothing, even remotely comes close. Emacs is not a text editor - it's a text orchestrator, you can control any text-relevant aspect in its computational vicinity - anything you see on your screen, local and remote computers.
Anything else might feel great for coding, but what do you when you need etymology or definition lookup, thesaurus exploration or translation in the middle of typing a comment? Well, you might say: "I'd bring the browser", but that's the equivalent of keeping your cup in the kitchen and whenever you need a sip of coffee, having to walk over there and back to your home office.
I can start conversing with a model virtually anytime, anywhere in Emacs. I can ask it to find me a command while I'm browsing the M-x menu, can you do that in Zed's command palette?
I can send a query asking for... I dunno, to fact check a sentence in the middle of typing it. While working with a model in ECA (Editor Coding Assistant) on some codebase, I can ask gptel (unrelated package) to correct my own prompt, using a different model, with dissimilar parameters. Or ask it to find me a way to summarize my current investigation, or combine multiple chats into a single pdf document, etc. The only limitation is your own imagination.
Not really, no. Zed’s a fine editor and I have nothing but good to say about it. I just happen to prefer Emacs more. Its keystrokes are burned into my muscle memory. And because its config files are literally executable code, there’s nothing too crazy for Emacs. Can it do XYZ? Absolutely, if you tell it to!
As to AI: Emacs continues to have zero built in. If you like it, though, it supports any agent you might want to use. I was highly amused the first time I wanted to use my own company’s agent and found it was already well-supported by a popular package.
And honestly, Emacs + a coding agent = magic for anyone who wanted to use it but was intimidated by Elisp. “Hey Droid, tell Emacs to swap the command and option keys on my Mac.” “Ok, done!” It’s a ludicrously powerful combo.
Way more than that if you count all the unpublished homebrew balls of glue.
I've been playing with writing my own lately. I just wrote a cool (IMHO) macro for defining tools, a defun-like form that at once generates both a callable Elisp function, and all the RPC pageantry needed for AI to call that function. From the same syntactic form.
The first tool I wrote was a function for retrieving source definitions of Elisp functions. Just enough to bootstrap the LLM harness into studying its own source code, reading one function's source and requesting new functions in turn, following the callgraph. It's neat to watch. It can dive into Emacs' internals too, since those are Elisp also.
That's correct. But most of the time when I see that question lately, it's because people want to be sure the editor doesn't have "built-in AI", whatever that might happen to mean. For example, I've seen lots of comments like "I like Zed but wish there was a version with no AI support in it whatsoever". I don't share that opinion. It's one I've seen many times.
I don't think that Emacs ships with any AI agent support today, and that makes some people happy, so I mentioned it. Note that I also mentioned that there are packages that do add that support, and I happily use them.
Personal opinions. How is it "top-notch", what does that even mean in practice?
I can tell you why working in Emacs feels empowering - when I'm working with an LLM, my Jira tickets are "transparent" - Emacs knows that any pattern that looks like 'XYZ-124354' is a Jira ticket - I can check the status, description, and other things on the fly. I don't need to switch to any other tool, need no browsing of the ticket. I can quickly convert it to a markdown link with a description (if I need to talk to my colleague about it). Same thing about 'foo/bar#42' pattern - Emacs recognizes it as a bug-reference token - it's an issue or a PR - I can immediately retrieve any info about it in the middle of typing.
I can collapse any large parts of my investigation without affecting my context and I can peek into them without expanding if I don't have to. I can change my freshly typed prompt using a different model, with different parameters, asking it I don't know, fact-check my assumptions or fill the gaps - all that while I'm typing it in my coding AI assistant buffer.
I can search for and pick a Slack thread interactively, without having to even switch to Slack, and tell the LLM to look into it, or simply extract some text or a snippet to paste it verbatim, or with modifications.
I can interactively pick a tab in my browser and insert its url, or grab a text from it. Without ever switching to the browser - it could even be not visible on my screen. I similarly can grab anything from my browser history.
Moreover, I can ask an LLM to change any aspect of my editor on the fly. If there's a use case I stumble on, I'd just slop the shit out of it and watch the agent drive my active Emacs instance, trying different things, popping up windows, taking screenshots and comparing results. I can name countless real, practical, everyday examples where I changed my workflow without missing a beat - some are trivial and some are more complex, and Emacs delivered. Without a hitch or a restart.
Emacs remains the absolute king for text manipulation of all kinds and text in the age of LLMs is simply everything. Nothing can beat Emacs when it comes to text, and oh boy have I tried so many other things.
Comments
Are there any Zed features you're going to miss? Is the AI notable?
Working with LLMs in Emacs is pure joy. LLMs are all about text and Emacs is the absolute king here. Nothing, even remotely comes close. Emacs is not a text editor - it's a text orchestrator, you can control any text-relevant aspect in its computational vicinity - anything you see on your screen, local and remote computers.
Anything else might feel great for coding, but what do you when you need etymology or definition lookup, thesaurus exploration or translation in the middle of typing a comment? Well, you might say: "I'd bring the browser", but that's the equivalent of keeping your cup in the kitchen and whenever you need a sip of coffee, having to walk over there and back to your home office.
I can start conversing with a model virtually anytime, anywhere in Emacs. I can ask it to find me a command while I'm browsing the M-x menu, can you do that in Zed's command palette?
I can send a query asking for... I dunno, to fact check a sentence in the middle of typing it. While working with a model in ECA (Editor Coding Assistant) on some codebase, I can ask gptel (unrelated package) to correct my own prompt, using a different model, with dissimilar parameters. Or ask it to find me a way to summarize my current investigation, or combine multiple chats into a single pdf document, etc. The only limitation is your own imagination.
Not really, no. Zed’s a fine editor and I have nothing but good to say about it. I just happen to prefer Emacs more. Its keystrokes are burned into my muscle memory. And because its config files are literally executable code, there’s nothing too crazy for Emacs. Can it do XYZ? Absolutely, if you tell it to!
As to AI: Emacs continues to have zero built in. If you like it, though, it supports any agent you might want to use. I was highly amused the first time I wanted to use my own company’s agent and found it was already well-supported by a popular package.
And honestly, Emacs + a coding agent = magic for anyone who wanted to use it but was intimidated by Elisp. “Hey Droid, tell Emacs to swap the command and option keys on my Mac.” “Ok, done!” It’s a ludicrously powerful combo.
Emacs has 35+ different AI tools.
https://www.reddit.com/r/emacs/comments/1uwm3c0/the_state_of...
Not being built-in into the core, means nothing. Emacs just doesn't operate that way.
Way more than that if you count all the unpublished homebrew balls of glue.
I've been playing with writing my own lately. I just wrote a cool (IMHO) macro for defining tools, a defun-like form that at once generates both a callable Elisp function, and all the RPC pageantry needed for AI to call that function. From the same syntactic form.
The first tool I wrote was a function for retrieving source definitions of Elisp functions. Just enough to bootstrap the LLM harness into studying its own source code, reading one function's source and requesting new functions in turn, following the callgraph. It's neat to watch. It can dive into Emacs' internals too, since those are Elisp also.
It's a fun project.
That's correct. But most of the time when I see that question lately, it's because people want to be sure the editor doesn't have "built-in AI", whatever that might happen to mean. For example, I've seen lots of comments like "I like Zed but wish there was a version with no AI support in it whatsoever". I don't share that opinion. It's one I've seen many times.
I don't think that Emacs ships with any AI agent support today, and that makes some people happy, so I mentioned it. Note that I also mentioned that there are packages that do add that support, and I happily use them.
Obviously, only a Mac user would need to do that. Every other person will just do this customization the normal way.
I went full circle on AI. For an in-editor experience, Zed is top notch. But I've found I actually prefer just to go back to claude code + Zed.
Personal opinions. How is it "top-notch", what does that even mean in practice?
I can tell you why working in Emacs feels empowering - when I'm working with an LLM, my Jira tickets are "transparent" - Emacs knows that any pattern that looks like 'XYZ-124354' is a Jira ticket - I can check the status, description, and other things on the fly. I don't need to switch to any other tool, need no browsing of the ticket. I can quickly convert it to a markdown link with a description (if I need to talk to my colleague about it). Same thing about 'foo/bar#42' pattern - Emacs recognizes it as a bug-reference token - it's an issue or a PR - I can immediately retrieve any info about it in the middle of typing.
I can collapse any large parts of my investigation without affecting my context and I can peek into them without expanding if I don't have to. I can change my freshly typed prompt using a different model, with different parameters, asking it I don't know, fact-check my assumptions or fill the gaps - all that while I'm typing it in my coding AI assistant buffer.
I can search for and pick a Slack thread interactively, without having to even switch to Slack, and tell the LLM to look into it, or simply extract some text or a snippet to paste it verbatim, or with modifications.
I can interactively pick a tab in my browser and insert its url, or grab a text from it. Without ever switching to the browser - it could even be not visible on my screen. I similarly can grab anything from my browser history.
Moreover, I can ask an LLM to change any aspect of my editor on the fly. If there's a use case I stumble on, I'd just slop the shit out of it and watch the agent drive my active Emacs instance, trying different things, popping up windows, taking screenshots and comparing results. I can name countless real, practical, everyday examples where I changed my workflow without missing a beat - some are trivial and some are more complex, and Emacs delivered. Without a hitch or a restart.
Emacs remains the absolute king for text manipulation of all kinds and text in the age of LLMs is simply everything. Nothing can beat Emacs when it comes to text, and oh boy have I tried so many other things.
And if the correct keystrokes are burned into your muscle memory, you can opt for Evil mode.