Claude Code Adds Cross-Session Messaging on macOS

[Keymaster]

Claude takes about 5x longer than almost any other AI. While it does well when working with very large documents, it does not do well at most other things. I'm just referring to the chat prompt, not Code or any kind of agents.

I also found that many times it takes it upon itself to do a ton of tasks that you never asked for...resulting in token usage. A couple days back I was trying to edit a 60-page bylaws document. Just by asking a question, it would redraft the entire document. After about 3 questions, it told me I was at my limit and had to wait 2 hours before I could continue.

This is a problem with AI, where it will rework a document from scratch just to fix an issue. I'm using Claude for coding, and I see others running into just this issue. I write code for astronomical instruments, and we have a scientist who is constantly creating new code using Claude Code, and it's clear that all he does is give a prompt and then blindly accept whatever comes out. That includes creating a whole new thing (which was then forced on the team for that instrument) that was 6000+ lines long, and every time he had Claude fix a bug it ended up revising the majority of that code...which meant the actual software engineers who were working on that instrument kept having 6000 line bombs dropped on them that they were told to make work. For those of you who aren't coders, just understand that's a huge workload, especially if you're told to find a minor bug somewhere in the mishmash of code (he didn't prompt it to write actual good, documented code either).

Eventually, I helped my office mate to get past the insanity that was giving him a mental block, and he coded up a real solution that fit into the instrument architecture, was 300 lines long, and he used Claude to do it to make the point of how you're actually supposed to use Claude as a coding tool. That hasn't stopped everyone else from continuing to do that...the same scientist just dropped a whole new GUI to automate the instrument onto everyone, and it's apparently the same nightmare again.

For myself, I have been using Claude to debug and find issues, and every time I make sure to tell it to figure the bug out first, explain what to do, and limit its work to where the bug is and not make changes anywhere else. I have also let it make its own choices and decisions to test what it can come up with, but even when I let it go it respects the boundaries I have set up for it. The key really is to put a fence around what you want it to do, give it clear instructions, and if you're having it make changes then make it clear to only change what you're looking at and document what has changed in the document. Do things that way, and it's actually pretty useful and doesn't run wild. This is also generally good coding practice anyways, small changes at a time, so enforcing that on Claude is pretty important if you want to turn out good code.

 

[chelsel]

This is 100% not true in any meaningful sense.

It’s possible the messaging interface is local, but in order for Claude to read the message it is immediately processed by Anthropic’s servers. That’s just how LLMs work.

You are correct. Note, Message passing could be a great way to coordinate sessions on different machines running local LLMs Qwen 3.6, via the Claude code interface. Utilizing multiple smaller machines in parallel.

 

[syklee26]

What is more essential than genmoji in macos messaging?

 

[bsolar]

This is 100% not true in any meaningful sense.

It’s possible the messaging interface is local, but in order for Claude to read the message it is immediately processed by Anthropic’s servers. That’s just how LLMs work.

Note that Claude Code can be configured to use different LLMs than the defaults from Anthropic's servers. E.g. this command starts Claude Code and makes it use a local LLM via ollama:

ollama launch claude --model qwen3.6

Capabilities may vary, but my understanding is that the cross-session-messaging functionality is provided by agentic tools, so a local LLM able to properly interact with the relevant tools should allow for completely offline use of the messaging feature.

PS: If the goal is to prevent potential traffic against Anthropic's servers, setting a couple additional configuration flags might be of interest:

CLAUDE_CODE_DISABLE_NONESSENTIAL_TRAFFIC=1
DISABLE_TELEMETRY=1

 

[jazzwill]

yippee more useless ai features !

Wondering do you:

1. Actually use AI to code, or
2. Just love sh*ting on everything that YOU think is useless?

 

[sevoneone]

Eventually, I helped my office mate to get past the insanity that was giving him a mental block, and he coded up a real solution that fit into the instrument architecture, was 300 lines long, and he used Claude to do it to make the point of how you're actually supposed to use Claude as a coding tool. That hasn't stopped everyone else from continuing to do that...the same scientist just dropped a whole new GUI to automate the instrument onto everyone, and it's apparently the same nightmare again.

I've been trying to impress on people I work with that the real gift of AI coding is not having a new tool or app overnight... the real gift is the ability to say "No" more and make something really refined in 2 weeks to a month that would have take 6 months to a year and a huge team before.

Everyone is way too eager to jump up up and down about slop they one-shot and have no brain about the fact that 50% of the infrastructure the app spun up is totally redundant, doesn't work with existing systems, or will become a dead end if/when it needs to scale.

 

[apple4everOG]

yippee more useless ai features !

Even better, we get to pay for it with massively more expensive memory!

 

[gleepskip]

Even better, we get to pay for it with massively more expensive memory!

It's not fuel or food. You can go without it if the price is too high.

 

[Audentia]

This is great news for engineers using claude code and other AI agents like this to build software. I hope an open standard emerges so that any agent can message each other, e.g. codex, claude code, grok build, local hosted agents and models (ollama etc.) can all talk to each other. Will solve a lot of annoying problems working with multiple agents at the same time.

 

[bluecoast]

I’m very much looking forward to multiple instances of Claude co-operating to hack into all of my accounts and stealing my data.

 

[apple4everOG]

It's not fuel or food. You can go without it if the price is too high.

It would be better though if the price was reasonable and not driven up by slop.

 

[mikebenton]

(1) - Say “don’t redraft the doc, just answer the question”
(2) - Consider upgrading to a paid tier for more usage capacity

I use the paid version. I use AI extensively. I appreciate the AI-101. This is what I run into with Claude every.single.day.

 

[nicholas.boncardo]

But are the agents doing things that people actually want? I really don't appreciate having AI in my gmail thank you very much.


Oh, is that what you meant by "running in the background"? While it's true that e.g. Netflix engineers are likely using AI agents to develop the Netflix client, that's not quite the same thing—when I click around the Netflix UI I'm not interacting with an AI agent at all.

Yes, you definitely are! All the suggestions you get in Netflix, the sort order in the categories which is determined by your usage data and a lot more are managed by AI.

 

[Wowfunhappy]

Yes, you definitely are! All the suggestions you get in Netflix, the sort order in the categories which is determined by your usage data and a lot more are managed by AI.

Okay, yes I guess you could call that an AI, although only because the term is meaningless, it can mean anything from your opponent in Mario Kart to ChatGPT.

I really don't think it would be correct to call your Netflix recommendations the result of an "AI agent" though, which is what the original post I was responding to said.