Cool idea and line of thought, obviously rough and early but it gets you thinking. “Software as clay” is obviously where the industry is heading, and as you say we’re approaching this from multiple angles… applying it directly in the browser is certainly an intriguing idea.
Why’d you make the prototype a separate browser instead of implementing with a chrome extension? Something like greasemonkey but with an LLM generating the scripts on the fly..
These drive my password manager nuts, especially when there are actually different logins for them. I just put a note in it saying exactly what service it's for.
What the hell is Kevin Scott (Microslop's CTO) doing with his time? How can any reasonable leader look at this disaster and go "hmm, yes, this looks like a sane and sustainable setup for future growth"?
They do, and those models are served by Microsoft. You pay a premium per “request” (what that means is not fully clear to me) for certain models. If you use the native chat extension in VSCode for GitHub CoPilot, with Opus model selected, you are not paying Anthropic. This counts against your GitHub Copilot subscription.
The Claude Code extension for VSCode from Anthropic will use your Claude subscription. But honestly it’s not very good - I use it but only to “open in terminal” (this adds some small quality of life features like awareness it’s in VSC so it opens files in the editor pane next to it).
This is my biggest frustration as a full time .NET developer. Its especially worse when you're searching for Visual Studio (IDE) specifics, and get results for VS Code. It bewilders me why a company that owns a search engine names their products so poorly.
I don’t use windows, so most of this doesn’t affect me, but I do use GitHub and VSCode. Can anyone clarify, once and for all, whether “GitHub Copilot” and “VSCode Copilot” (sic?) are the same product? The documentation isn’t even clear, and it’s important because it affects billing. How do these two products interact and where do they NOT overlap?
This confusion even bleeds into other coding harnesses. I have no idea which GitHub MCP server I setup in Claude Code, but the domain has “githubcopilot” in it. Am I burning copilot tokens (or “requests” or whatever is their billing unit) when I use this from Claude?
Git is a distributed source control system. It's open source and you can use it to version source code on your drive and/or a remote git repository.
Github is one of the most popular git repository hosts. In addition to source repositories, it has other services like issue tracking and wikis.
A while back, Microsoft bought Github.
"Github Copilot" is a service you can buy (with limited free sku) from Github that adds AI capabilities to your Github subscription.
One of the ways you can use Github Copilot is by using the GitHub Copilot extension for VSCode. This extension lets you use chat inside VSCode in such a way that it can read and write code. It lets you pick which LLM model you want to use: Claude Sonnet, Opus, OpenAI GPT, etc., from the ones they support.
Note you don't need another subscription if you only use Github Copilot. They pay Anthropic, you pay Github. You _might_ want another subscription directly with Anthropic if, say, you want to use Claude Code instead.
"VSCode Copilot" isn't a thing. Some people might call Github Copilot extension for VSCode "VSCode Copilot".
Github MCP server lets AI tools like GitHub Copilot extension for VSCode, Claude Code, or any tool that supports MCP use your Github account to do things like pull requests, read issues, etc. Just using it from Claude Code would not use Github Copilot tokens, UNLESS you used it to work against your Github Copilot service. You would not need a Github Copilot subscription to use it for example to create a pull request or read an issue. But it would use your Github Copilot tokens if, say, you used the MCP from Claude Code to assign a task to Github Copilot. It uses githubcopilot domain because they built it mostly for Github Copilot to use, though MCP is an open standard so it can be used from any MCP-supporting AI tool.
It's massively cheaper. Copilot charges per request, which with some clever prompting, can lead to huge amounts of work being done at fractions of the cost of Claude Code. Millions of tokens for mere pennies. MS must be taking a huge hit somewhere, because I'm probably getting 10-20x my value out of GH relative to CC.
I am not locked in to Anthropic, either. I can easily switch between GPT and Gemini models based on how I think each would perform in various scenarios. That's a big win. I use a lot of design with Opus, implement with GPT 5.4.
Also, Github Copilot CLI is pretty much at feature parity (for the stuff that matters) with Claude Code. Using both at work and home, I don't think there's much difference in features between the two. Maybe I'm not a super power user, and just a regular dumb user, but GH doesn't seem buggy and everything I think I'd want to do with CC I can do with GH.
I'm spending a literal fortune on CC - we also have GH Copilot but the devs imply that CC is better? Will the Github Copilot let us access skills and agent frameworks in CC?
Devs say a lot of uninformed things. With a heavy predisposition to hating the "legacy" monoliths that are Microsoft and by association GitHub.
Yes, Copilot supports skills. Practically all agents support very similar feature sets or are actively building up parity support if not already there. The only real difference between systems is the prompt and payment method. Copilot even allows you to use Anthropic's own skills repository: https://github.com/anthropics/skills
It's a bit rich to go around calling people uninformed because they prefer one harness to another, particularly when you are recommending GHC as comparable to CC.
It’s a bit rich to take the most negative interpretation of my statement, and moreso telling of your insecurities that you chose to be so offended.
And, ultimately, proving my point. Did you actually explain why you thought it’s superior? Or is it just because GitHub bad? Have you even tried it recently?
IME is is less capable of performing complex work, more frequently goes down blind alleys and needs correcting, that kind of thing. It's night and day vs CC.
Make it write a skill and rule hook for PreCompact to do a handoff that explains what was worked on, what to know, and what to do next. If it goes off the rails after compaction then it won’t be great in a new session either, and you want to make sure you maximize continuity or development will be unsustainable. A backlog.md and improvements.md workflow also helps with this (ticket numbers, descriptions, “focus on BACK-0075,” etc.)
And this has been comparing like for like with CC - say Opus 4.6 on the same reasoning effort? Hasn’t been my experience particularly but fair enough. I do tend to use them in different situations (CC outside of work).
Even if it is close, maybe GHC CLI has improved in the last month since I last used it, I know you didn't say it but calling people uninformed because they prefer one or the other is just wrong.
I’d agree, though maybe there’s a more charitable reading of the OP - “uninformed” is one of those accusations that it’s rarely very polite or fair to level against an individual but sometimes is reasonable against a group based on observation. My experience would be that it’s true that “devs says lots of uninformed things” - and I’d include myself in that. It’s been my experience that it’s particularly tough in this space at this time because:
1. Tooling is changing very fast but people tend to form sticky opinions (reasonably enough - there’s only so much time in the world).
2. It’s just hard to form robust objective opinions - you have to make a real effort to build test cases and evaluation processes and generally the barrier to entry there is pretty high.
So - I agree, calling people uninformed is not a great way to win them over, but maybe that’s the price of living in a world of anecdotes which become fixed in people’s minds.
Claude (and most other models) in GitHub Copilot still only have 200k context, with a hefty amount being reserved for some reason. It's 1M at many other providers.
Do hacks like “read prompt.md, and follow its instructions. When you’re done, read it again and follow its instructions.” And then you have some background process appending to the file to keep it warm and you just keep writing there?
You could do that. I was just trying to say that if you make your original prompt complete enough, and you have well-defined success criteria, you can tell it to keep going until they are met.
> "Fix the following compile errors" -> one shot try and stops.
> "Fix the following compile errors. When done, test your work and continue iterating until build passes without error" -> same cost but it gets the job done.
There is a limit on how much copilot can do in one request, pretty generous but after some time vscode will say "this request is taking very long, do you want to continue" and that would count as a seperate request
I use it because they offer absurdly cheap prices that they're clearly losing money on. I can get $1000 at API prices of Opus 4.6, for in the range of $2 my cost through copilot.
Tighter (read better) integration with VSCode and Github than what you could get running claude code on the side.
Your question does raise a valid point - Github Copilot's value proposition is fairly limited in my opinion. Not to say worthless but limited and clearly varies depending on how Githubbey your dev workflows are.
Yeah, the workflow is superb. That’s what I miss most using Claude in a terminal inside VSCode. It doesn’t integrate with VSCode native diff tools like the native VSCode (GitHub Copilot does. The Claude extension in non-terminal mode is crap.
From a user point of view there's no real reason for it, from an admin point of view if your team is already using Github Enterprise then deploying it is basically hitting a toggle switch, and it has some more fine grained controls about what it can or can't do compared to Claude Code.
you can also get a service contract via MS quite easily/cheaply, which mightnot help you with hard problems but does solve the easy ones. example: in earlydays we bought OpenAI API directly and via Azure; when we needed account service we got it immediately from MS instead of waitlists from OpenAI.
> I don’t use windows, so most of this doesn’t affect me, but I do use GitHub and VSCode. Can anyone clarify, once and for all, whether “GitHub Copilot” and “VSCode Copilot” (sic?) are the same product? The documentation isn’t even clear, and it’s important because it affects billing. How do these two products interact and where do they NOT overlap?
There is no VSCode Copilot. There is Github Copilot integration inside VS Code.
Tldr yes they're the same agentic harness in different UIs. Web browser, android app, ide extension, cli tool. They all change the "how" but not the "what".
You buy premium you get more prompts and models.
This is an excellent post and great reference material. I’ve done this a few times before and the information was scattered all over the place. I appreciate the clear and concise writing here. I even added it to my HN favorites - a rare accolade!
One thing I’d add, is that the best explanation I’ve ever seen for this, is the famous diagram [0] on Wikipedia of the netfilter API — I remember when I saw that, everything clicked into place. I’m not sure how up to date it is now, but it’s really good.
I’m surprised people are still using OpenClaw. I assumed they’d have switched to Nanoclaw or Nemoclaw. Is OpenClaw just that much better, or is it all inertia?
None of them, but prefer ones written with engineering rigor and security in mind. Having an unvetted plugin ecosystem with code that runs unsandboxed is laughably naive
We use Pi at work (where we pay per token) and I’d love to use it personally too. From what I’ve read, nobody has been banned for using Pi yet… I wonder if Anthropic minds this much as long as it’s still human usage, or if they’re mostly focused on stamping out the autonomous harnesses. Unfortunately Pi is also what OpenClaw uses so it could easily get swept up in the enforcement attention.
Or maybe I’ll just get a Codex subscription instead. OpenAI has semi-officially blessed usage of third party harnesses, right?
It appears that OpenAI has blessed third party harnesses. I know they officially support OpenCode and they have this on their developer portal:
"Developers should code in the tools they prefer, whether that's Codex, OpenCode, Cline, pi, OpenClaw, or something else, and this program supports that work."
Obviously, the context is that OpenAI is telling open source developers who are using free subscriptions/tokens from the Codex for Open Source program that they can use any harness they want. But it would be strange for that to not extend to paying subscribers.
They have, but they also just announced this week that for business and enterprise plans, they’re switching from quotas for codex to token use based pricing, and I would expect that to eventually propagate to all their plans for all the same reasons.
I’d be surprised if that propagated to personal subscription plans, simply because it would put them at a huge competitive disadvantage against Anthropic, which they’ve already signaled they care about by saying they allow third-party harnesses. But I wouldn’t be surprised if they required third-party harnesses to use per-token billing, since that’d put them on par with Anthropic.
I've done this with Cursor because I have similar issues with inconsistent allowance consumption there. I mostly use Claude models but I've had to disable Opus 4.6 because it just EATS tokens in it's thinking steps.
Why’d you make the prototype a separate browser instead of implementing with a chrome extension? Something like greasemonkey but with an LLM generating the scripts on the fly..
reply