I like how in the demo video there's a squiggle emphasis on Claude's "Good Idea!" in response to a user clarification, when it's more common among vibe coders that that less glazing is better and they just want the LLM to write code.
Pair programming is still one of the best ways to knowledge transfer between two programmers in a high throughput manner. Humans learn by doing, building synaptic connections.
I wonder if a shared Claude Code instance has the same effect?
I haven't seen anyone talking about using agents in this way. I wonder if it would be helpful for learning e.g. a new language or codebase to have the human write the code while the agent takes the role of the "backseat driver" in the pair programming dynamic.
The most interesting parts of this to me are somewhat buried:
- Claude Code has been added to iOS
- Claude Code on the Web allows for seamless switching to Claude Code CLI
- They have open sourced an OS-native sandboxing system which limits file system and network access _without_ needing containers
However, I find the emphasis on limiting the outbound network access somewhat puzzling because the allowlists invariably include domains like gist.github.com and dozens of others which act effectively as public CMS’es and would still permit exfiltration with just a bit of extra effort.
I’m wondering if it would be possible to use the new skills feature or agents with this. Without the agents or the skills, I don’t know how useful this would be.
No relations to them, but I've started using Happy[0]'s iOS app to start and continue Claude Code sessions on my iPhone. It allows me to run sessions on a custom environment, like a machine with a GPU to train models
Very curious to see what usage limits are like for paid plans. Anthropic was already experiencing issues with high-volume model usage for Pro and Max users. I hope their infrastructure is able to adequately support running these additional coding environments on top of model inference.
Just to be clear, I'm excited for the capability to use Claude Code entirely within the browser. However, I've heard reports of Max users experiencing throttled usage limits in recent months, and am concerned as to whether this will exacerbate that issue or not.
I'm using Claude Code locally a lot, occasionally with a couple parallel session.
I was very happy when they made the GitHub Action - I used it quite a bit, but in practice I got frustrated that I effectively only get a single back-and-forth out of it, I can't really "continue the conversation without losing context" - Sure, I can respond to it in the PR it makes, but that will be a fresh session with a fresh empty context.
So, as much as I don't like moving out of my standard development workflow with my tools, I think this could be quite useful. The ability to interrupt and/or continue a conversation should be very nice.
My main worry is - usually my unit tests and integration tests rely on a postgres database running on the machine, and it's not obvious to me if I can spin that up here?
I've been using Happy Coder[0] for some time now on web and mobile. I run it `--yolo` mode on an isolated VM across multiple projects.
With Happy, I managed to turn one of these Claude Code instances into a replacement for Claude that has all the MCP goodness I could ever want and more.
It's really solid. It's effectively a web (and native mobile) UI over Claude Code CLI, more specifically "claude --dangerously-skip-permissions".
Anthropic have recognized that Claude Code where you don't have to approve every step is massively more productive and interesting than the default, so it's worth investing a lot of resources in sandboxing.
> it's worth investing a lot of resources in sandboxing.
I tend to agree. There’s an opportunity to make it easy to have Claude be able to test out workflows/software within Debian, RPM, Windows, etc… container and VM sandboxes. This could be helpful for users that want to release code on multiple platforms and help their own training and testing, which they seem to be heavily invested in given all the “How Am I doing?” prompts we’re getting.
great points @simonw - I, incredibly, haven't ever tried --dangerously-skip-permissions yet for any "real" projects. I generally find that it stops itself for good reason.
I got so used to having Claude Code read some of my MCP tools, and was bummed to see that it couldn't connect to them yet on the web.
Pretty cool though! Will need to use it for some more isolated work/code edits. Claude Code is now my workhorse for a ton of stuff including non-coding work (esp. with the right MCPs)
We were heavy users of Claude Code ($70K+ spend per year) and have almost completely switched to codex CLI. I'm doing massive lifts with it on software that would never before have been feasible for me personally, or any team I've ever run. I'll use Claude Code maybe once every two weeks as a second set of eyes to inspect code and document a bug, with mixed success. But my experience has been that initially Claude Code was amazing and a "just take my frikkin money" product. Then Codex overtook CC and is much better at longer runs on hard problems. I've seen Claude Code literally just give up on a hard problem and tell me to buy something off the shelf. Whereas Codex's ability to profoundly increase the capabilities of a software org is a secret that's slowly getting out.
I don't have any relationship with any AI company, and honestly I was rooting for Anthropic, but Codex CLI is just way way better.
Also Codex CLI is cheaper than Claude Code.
I think Anthropic are going to have to somehow leapfrog OpenAI to regain the position they were in around June of this year. But right now they're being handed their hat.
As we all know here, if the the title of this post was about Codex on the web, the top comment would have been about using Claude instead.
YMMV, but this definitely doesn't track with everything I've been seeing and hearing, which is that Codex is inferior to Claude on almost every measure.
Im not saying this is a paid endorsement but the internet is dead and I wonder what openAI would pay, if they could, to get such a glowing review as top comment on HN
> I've seen Claude Code literally just give up on a hard problem and tell me to buy something off the shelf
I've been seeing more of this lately despite initial excellent results. Not sure what's going on, but the value is certainly dropping for me. I'll have to check out codex. CLI integration is critical for me at this point. For me it is the only thing that actually helps realize the benefits of LLM models we have today. My last NixOS install was completely managed by Claude Code and it worked very well. This was the result of my latest frustrations:
Though I know the statement it made isn't "true". I've had much better luck pursuing other implementation paths with CC in the same space. I could have prompted around this and should have reset the context much earlier but I was drunk "coding" at that point and drove it into a corner.
My experience is that if I know what I want, CC will produce better code, given I specify it correctly. The planning mode is great for this too, as we can "brainstorm" and what I have seen help a lot is if I ask questions about why it did a certain way. Often it'll figure out on its own why that's wrong, but sometimes it requires a bit of course correction.
On the other hand, last time I tried GPT-5 from Cursor, it was so disappointing. It kept getting confused while we were iterating on a plan, and I had to explain to it multiple times that it's thinking about the problem the same way. After a while I gave up, opened a new chat and gave it my own summary of the conversation (with the wrong parts removed) and then it worked fine. Maybe my initial prompt was vague, but it continually seemed to forget course corrections in that chat.
I mostly tend to use them more to save me from typing, rather than asking it to design things. Occasionally we do a more open ended discussion, but those have great variance. It seems to do better with such discussions online than within the coding tool (I've bounced maths/implementation ideas off of while writing shaders on a personal project)
> We were heavy users of Claude Code ($70K+ spend per year)
Claude code has only been generally available since May last year (a year and half ago)... I'm surprised by the process that you are implying; within a year and a half, you both spent 70k on claude code, and knew enough about it and its competition to switch away from it? I dont think I'd be able to due diligence even if LLM evaluation was my fulltime job. Let alone the fact that the capabilities of each provider are changing dramatically every few weeks.
Feels like with every announcement there’s the same comment: “this LLM tool I’m using now is the real deal, the thing I was using previously and spending stupid amounts of money on looked good but failed at XYZ, this new thing is where it’s at”. Rinse and repeat.
Which means it wasn’t true any of the previous times, so why would it be true this time? It feels like an endless loop of the “friendship ended” meme with AI companies.
It’s much more likely commenters are still in the honeymoon hype phase and (again) haven’t found the problems because they’re hyper focused on what the new thing is good at that the previous one wasn’t, ignoring the other flaws. I see that a lot with human relationships as well, where people latch on to new partners because they obviously don’t have the big problem that was a strain on the previous relationship. But eventually something else arises. Rinse and repeat.
When you say Claude Code, what model do you refer to? CC with Opus still outperforms Codex (gpt-5-codex) for me for anything I do (Rust, computer graphics-related).
However, Anthropic restricted Opus use for Max plan users 10 days or so ago severly (12-fold from 40h/week down to 5h week) [1].
Sonnet is a vastly inferioir model for my use cases (but still frequently writes better Rust code than Codex). So now I use Codex for planning and Sonnet for writing the code. However, I usually need about 3--5 loops with Codex reviewing, Sonnet fixing, rinse & repeat.
Before I could use one-shot Opus and review myself directly, and do one polish run following my review (also via Opus). That was possible from June--mid October but no more.
Costs are 6x cheaper and it's way faster and good at test writing and tool calling. It some times can be a bit messy though so use Gemini or Claude or codex for that hard problems....
In agreement. Large caveats that can explain differing opinions (that I've experienced) are:
* Is really only magic on Linux or WSL. Mediocre on Windows
* Is quite mediocre at UI code but exceptional at backend, engineering, ops, etc. (I use Claude to spruce up everything user facing -- Codex _can_ mirror designs already in place fairly well).
* Exceptional at certain languages, OK at others.
* GPT-5 and GPT-5-Codex are not the same. Both are models used by the Codex CLI and the GPT-5-Codex model is recent and fantastically good.
* Codex CLI is not "conversational" in the way that Claude is. You kind of interact with it differently.
I often wonder about the impact of different prompting styles. I think the WOW moment for me is that I am no longer returning to code to find tangled messes, duplicate silo'd versions of the same solution (in a different project in the same codebase), or strangely novice style coding and error handling.
As a developer for 20yrs+, using Codex running the GPT-5-Codex model has felt like working with a peer or near-peer for the first time ever. I've been able to move beyond smaller efforts and also make quite a lot of progress that didn't have to be undone/redone. I've used it for a solid month making phenomenal progress and able to offload as-if I had another developer.
Honestly, my biggest concern is that OpenAI is teasing this capable model and then pulls the rug in a month with an "update".
As for the topic at hand, I think Claude Code has without a doubt the best "harness" and interface. It's faster, painless, and has a very clean and readable way of laying out findings when troubleshooting. If there were a cheap and usable version of Opus... perhaps that would keep Claude Code on the cutting edge.
This was my experience too until a couple weeks ago, when Codex suddenly got dumbed down.
Initially, I had great success with codex medium- I could refactor with confidence, code generally ran on the first or second try, etc.
Then when that suddenly dumbed down to Claude Sonnet 3.5 quality I moved to GPT5 High to get back what had been lost. That was okay for a few days. Now GPT5 High has dropped to Claude Sonnet 3.5 quality.
We’re moving almost entirely to Codex, first because often it’s just better, and second because it’s much cheaper. It’s a bet that they’re better now, but given capacity and funding, they’ll be better later too.
The only edge Claude has is context window, which we do sometimes hit, but I’m sure that gap will close.
I really want this but for Azure Devops. If you're not familiar, Microsoft owns both Github and Azure Devops, and both do similar: git repos and project management. I can use Github Copilot, Claude Code CLI, etc. against code on my disk, including Azure Devops MCP. But what I can't easily do is like Github Copilot Agent and apparently this Claude Code on Web: Assign a ticket to @SomeAi and have a PR show up in a few minutes. Can't change to github for _reasons_.
Would love any suggestions if anyone in a similar story.
Nit about doing your AI interfaces on the Web: I really want claude.ai and chatgpt.com to offer a standard username+password login without 2FA. The kind my privacy-friendly browser of short-lived sessions can complete in a couple clicks, like for most other SaaSes, and then I'm in and using the tool.
I don't want to leak data either way by using some "let's throw SSO from a sketchy adtech company into the trust loop".
I don't want to wait a minute for Anthropic's login-by-email link, and have the process slam the brakes on my workflow and train of thought.
I don't want to wait a minute for OpenAI's MFA-by-email code (even though I disabled that in the account settings, it still did it).
I don't want to deal with desktop clients I don't trust, or that might not keep up with feature improvements. Nor have to kludge up a clumsy virtualization sandbox for an untrusted client, just to ask an LLM questions that could just be in a Web browser.
Does this work inside docker containers like Codex? Stuff like `testcontainers` is unusable with that architecture because you need access to docker itself.
Just played around with it the fact it’s on the phone is a big bonus.
I have setup a little workflow where given linear tags it sets up a work tree on my dev box installs deps and starts the implementation so I can take it over I prefer this workflow to the fully managed cloud based solutions.
This kind of fits in for issues where I’m basically sure I won’t have to take it over (and it can do it fully on its own). Which aren’t that many.
Very simple example there was a warning pop up on something where I thought there shouldn’t be now it’s done fully automatically from my phone in 5 mins. I quite like that these small changes become so easy.
I got my environment working well with Codex's Cloud Task. Trying to same repo with Claude Code Web (which started off with Claude Code CLI mind you), and the yarn install just hangs with no debuggable output.
102 comments
[ 1952 ms ] story [ 1044 ms ] threadI wonder if a shared Claude Code instance has the same effect?
- Claude Code has been added to iOS
- Claude Code on the Web allows for seamless switching to Claude Code CLI
- They have open sourced an OS-native sandboxing system which limits file system and network access _without_ needing containers
However, I find the emphasis on limiting the outbound network access somewhat puzzling because the allowlists invariably include domains like gist.github.com and dozens of others which act effectively as public CMS’es and would still permit exfiltration with just a bit of extra effort.
[0] https://github.com/slopus/happy/
Just to be clear, I'm excited for the capability to use Claude Code entirely within the browser. However, I've heard reports of Max users experiencing throttled usage limits in recent months, and am concerned as to whether this will exacerbate that issue or not.
I'm using Claude Code locally a lot, occasionally with a couple parallel session.
I was very happy when they made the GitHub Action - I used it quite a bit, but in practice I got frustrated that I effectively only get a single back-and-forth out of it, I can't really "continue the conversation without losing context" - Sure, I can respond to it in the PR it makes, but that will be a fresh session with a fresh empty context.
So, as much as I don't like moving out of my standard development workflow with my tools, I think this could be quite useful. The ability to interrupt and/or continue a conversation should be very nice.
My main worry is - usually my unit tests and integration tests rely on a postgres database running on the machine, and it's not obvious to me if I can spin that up here?
With Happy, I managed to turn one of these Claude Code instances into a replacement for Claude that has all the MCP goodness I could ever want and more.
[0]: https://happy.engineering/
without 200% price increase in 3 years, there's no way any of these AI companies will survive.
It's really solid. It's effectively a web (and native mobile) UI over Claude Code CLI, more specifically "claude --dangerously-skip-permissions".
Anthropic have recognized that Claude Code where you don't have to approve every step is massively more productive and interesting than the default, so it's worth investing a lot of resources in sandboxing.
I tend to agree. There’s an opportunity to make it easy to have Claude be able to test out workflows/software within Debian, RPM, Windows, etc… container and VM sandboxes. This could be helpful for users that want to release code on multiple platforms and help their own training and testing, which they seem to be heavily invested in given all the “How Am I doing?” prompts we’re getting.
Pretty cool though! Will need to use it for some more isolated work/code edits. Claude Code is now my workhorse for a ton of stuff including non-coding work (esp. with the right MCPs)
I don't have any relationship with any AI company, and honestly I was rooting for Anthropic, but Codex CLI is just way way better.
Also Codex CLI is cheaper than Claude Code.
I think Anthropic are going to have to somehow leapfrog OpenAI to regain the position they were in around June of this year. But right now they're being handed their hat.
YMMV, but this definitely doesn't track with everything I've been seeing and hearing, which is that Codex is inferior to Claude on almost every measure.
Fails to escalate permissions, gets derailed, loves changing too many things everywhere.
GPT5 is good, but codex is not.
I've been seeing more of this lately despite initial excellent results. Not sure what's going on, but the value is certainly dropping for me. I'll have to check out codex. CLI integration is critical for me at this point. For me it is the only thing that actually helps realize the benefits of LLM models we have today. My last NixOS install was completely managed by Claude Code and it worked very well. This was the result of my latest frustrations:
https://i.imgur.com/C4nykhA.png
Though I know the statement it made isn't "true". I've had much better luck pursuing other implementation paths with CC in the same space. I could have prompted around this and should have reset the context much earlier but I was drunk "coding" at that point and drove it into a corner.
I thought claude code is still better in tool calling and something like that
On the other hand, last time I tried GPT-5 from Cursor, it was so disappointing. It kept getting confused while we were iterating on a plan, and I had to explain to it multiple times that it's thinking about the problem the same way. After a while I gave up, opened a new chat and gave it my own summary of the conversation (with the wrong parts removed) and then it worked fine. Maybe my initial prompt was vague, but it continually seemed to forget course corrections in that chat.
I mostly tend to use them more to save me from typing, rather than asking it to design things. Occasionally we do a more open ended discussion, but those have great variance. It seems to do better with such discussions online than within the coding tool (I've bounced maths/implementation ideas off of while writing shaders on a personal project)
Claude code has only been generally available since May last year (a year and half ago)... I'm surprised by the process that you are implying; within a year and a half, you both spent 70k on claude code, and knew enough about it and its competition to switch away from it? I dont think I'd be able to due diligence even if LLM evaluation was my fulltime job. Let alone the fact that the capabilities of each provider are changing dramatically every few weeks.
Which means it wasn’t true any of the previous times, so why would it be true this time? It feels like an endless loop of the “friendship ended” meme with AI companies.
https://knowyourmeme.com/editorials/guides/what-is-the-frien...
It’s much more likely commenters are still in the honeymoon hype phase and (again) haven’t found the problems because they’re hyper focused on what the new thing is good at that the previous one wasn’t, ignoring the other flaws. I see that a lot with human relationships as well, where people latch on to new partners because they obviously don’t have the big problem that was a strain on the previous relationship. But eventually something else arises. Rinse and repeat.
However, Anthropic restricted Opus use for Max plan users 10 days or so ago severly (12-fold from 40h/week down to 5h week) [1].
Sonnet is a vastly inferioir model for my use cases (but still frequently writes better Rust code than Codex). So now I use Codex for planning and Sonnet for writing the code. However, I usually need about 3--5 loops with Codex reviewing, Sonnet fixing, rinse & repeat.
Before I could use one-shot Opus and review myself directly, and do one polish run following my review (also via Opus). That was possible from June--mid October but no more.
[1] https://github.com/anthropics/claude-code/issues/8449
Costs are 6x cheaper and it's way faster and good at test writing and tool calling. It some times can be a bit messy though so use Gemini or Claude or codex for that hard problems....
* Is really only magic on Linux or WSL. Mediocre on Windows
* Is quite mediocre at UI code but exceptional at backend, engineering, ops, etc. (I use Claude to spruce up everything user facing -- Codex _can_ mirror designs already in place fairly well).
* Exceptional at certain languages, OK at others.
* GPT-5 and GPT-5-Codex are not the same. Both are models used by the Codex CLI and the GPT-5-Codex model is recent and fantastically good.
* Codex CLI is not "conversational" in the way that Claude is. You kind of interact with it differently.
I often wonder about the impact of different prompting styles. I think the WOW moment for me is that I am no longer returning to code to find tangled messes, duplicate silo'd versions of the same solution (in a different project in the same codebase), or strangely novice style coding and error handling.
As a developer for 20yrs+, using Codex running the GPT-5-Codex model has felt like working with a peer or near-peer for the first time ever. I've been able to move beyond smaller efforts and also make quite a lot of progress that didn't have to be undone/redone. I've used it for a solid month making phenomenal progress and able to offload as-if I had another developer.
Honestly, my biggest concern is that OpenAI is teasing this capable model and then pulls the rug in a month with an "update".
As for the topic at hand, I think Claude Code has without a doubt the best "harness" and interface. It's faster, painless, and has a very clean and readable way of laying out findings when troubleshooting. If there were a cheap and usable version of Opus... perhaps that would keep Claude Code on the cutting edge.
Initially, I had great success with codex medium- I could refactor with confidence, code generally ran on the first or second try, etc.
Then when that suddenly dumbed down to Claude Sonnet 3.5 quality I moved to GPT5 High to get back what had been lost. That was okay for a few days. Now GPT5 High has dropped to Claude Sonnet 3.5 quality.
There's nothing left to fallback to.
The only edge Claude has is context window, which we do sometimes hit, but I’m sure that gap will close.
Would love any suggestions if anyone in a similar story.
https://developer.microsoft.com/blog/azure-devops-with-githu...
(if they want employee to use more AI, ditch ADO, embrace GitHub)
I don't want to leak data either way by using some "let's throw SSO from a sketchy adtech company into the trust loop".
I don't want to wait a minute for Anthropic's login-by-email link, and have the process slam the brakes on my workflow and train of thought.
I don't want to wait a minute for OpenAI's MFA-by-email code (even though I disabled that in the account settings, it still did it).
I don't want to deal with desktop clients I don't trust, or that might not keep up with feature improvements. Nor have to kludge up a clumsy virtualization sandbox for an untrusted client, just to ask an LLM questions that could just be in a Web browser.
I have setup a little workflow where given linear tags it sets up a work tree on my dev box installs deps and starts the implementation so I can take it over I prefer this workflow to the fully managed cloud based solutions.
This kind of fits in for issues where I’m basically sure I won’t have to take it over (and it can do it fully on its own). Which aren’t that many.
Very simple example there was a warning pop up on something where I thought there shouldn’t be now it’s done fully automatically from my phone in 5 mins. I quite like that these small changes become so easy.
I got my environment working well with Codex's Cloud Task. Trying to same repo with Claude Code Web (which started off with Claude Code CLI mind you), and the yarn install just hangs with no debuggable output.