I was thinking about doing this. If we treat ChatGPT like your local llm server, you basically open up a browser session, intercept all traffic and stick a local OpenAI compatible rest server that mutilates your input upstream through the ChatGPT ui, and then intercepts downstream and pushes it out your server.
API call to server
Then
Server leverages chrome plugin to send your llm call through the plugin , through the ui, wait for response , intercept, repackage and reroute it back out as a server response.
Then that $20 sub stretches to all uses cases for a developer.
That's a clever idea, and a different direction from where we are right now. Today it only goes one way, from chat into your coding agent.
The case I hit all the time is that an idea strikes on my commute, so I pull out my phone and start brainstorming in the Claude app. By the time I'm home I've done the research and chased down the tangents in that conversation. I open my laptop, use Skillsync to continue it in Claude Code, and start on the implementation.
How would they know? Only if you use a coding harness, but one could write their own harness with less obvious system prompts.
I’m not encouraging one do this. I mostly route between the free models on OpenRouter and OpenCode free models, and I mostly write simple apps where I can casually oversee even a small model and get awesome results.
But if you are out there spending 200 dollars a month, consider a few hacks, because that’s just overpriced. We are developers after all, they aren’t supposed to be able to trick us :)
If I were you I would look into maybe like analyzing sessions across harnesses? I’m not sure, some way to leverage this shared schema work that is not syncing
Because syncing is trivially solved by just telling the agent to keep a work journal in markdown
We do that, and it's really the point of the portability work. Once every harness's sessions live in one schema, you can manage and search all of them in one place.
The markdown journal is a fair way to do it, and a popular one. It has a couple of limits though.
A journal comes out of a prompt, so it records one way of looking at the session. Whatever that prompt didn't ask about is gone, and you usually find out which detail you needed much later. But the raw session still has it.
Being able to unify the schema lets you do things notes can't. You can continue a teammate's session on your own machine, or pick up a claude.ai conversation in Claude Code or Codex. The new agent gets the transcript: all the decisions in the trajectory, not a summary of it.
To say what the author said in another way- thats a garbage approach and completely misses the point. Sure, you can completely destroy the session by distilling it into a markdown- but thats like explaining what you did yesterday, vs having an immutable log of what you did yesterday.
This is good, I like it!! I've been mostly working with Claude and Chatgpt for separate stuff, however sometimes one of them is stuck and I could definitely use this to start a session with the other one and see if it figures it out.
Have you tested to see if there was any degradation happening when switching from one model to the other ? I mean some must be inevitable (maybe not!), but how much?
Nars here, one of the co-founders. With longer sessions, the receiving agent sometimes decides to compact it. But the whole transcript still exists and is accessible
I was thinking about doing this. If we treat ChatGPT like your local llm server, you basically open up a browser session, intercept all traffic and stick a local OpenAI compatible rest server that mutilates your input upstream through the ChatGPT ui, and then intercepts downstream and pushes it out your server.
API call to server
Then
Server leverages chrome plugin to send your llm call through the plugin , through the ui, wait for response , intercept, repackage and reroute it back out as a server response.
Then that $20 sub stretches to all uses cases for a developer.
That's a clever idea, and a different direction from where we are right now. Today it only goes one way, from chat into your coding agent.
The case I hit all the time is that an idea strikes on my commute, so I pull out my phone and start brainstorming in the Claude app. By the time I'm home I've done the research and chased down the tangents in that conversation. I open my laptop, use Skillsync to continue it in Claude Code, and start on the implementation.
Good way to be banned
How would they know? Only if you use a coding harness, but one could write their own harness with less obvious system prompts.
I’m not encouraging one do this. I mostly route between the free models on OpenRouter and OpenCode free models, and I mostly write simple apps where I can casually oversee even a small model and get awesome results.
But if you are out there spending 200 dollars a month, consider a few hacks, because that’s just overpriced. We are developers after all, they aren’t supposed to be able to trick us :)
isn't that what they did to opencode? when they were using the subs by reverse engineering their APIs
I have made simple macos app to export my claude sessions to markdown(plus ordinal jsonl) and i just continuously sync it to dropbox.
i do the same for apple notes and imessage. so very portable
https://apps.apple.com/us/app/exporter-notes-to-markdown/id6...
If I were you I would look into maybe like analyzing sessions across harnesses? I’m not sure, some way to leverage this shared schema work that is not syncing
Because syncing is trivially solved by just telling the agent to keep a work journal in markdown
We do that, and it's really the point of the portability work. Once every harness's sessions live in one schema, you can manage and search all of them in one place.
The markdown journal is a fair way to do it, and a popular one. It has a couple of limits though.
A journal comes out of a prompt, so it records one way of looking at the session. Whatever that prompt didn't ask about is gone, and you usually find out which detail you needed much later. But the raw session still has it.
Being able to unify the schema lets you do things notes can't. You can continue a teammate's session on your own machine, or pick up a claude.ai conversation in Claude Code or Codex. The new agent gets the transcript: all the decisions in the trajectory, not a summary of it.
To say what the author said in another way- thats a garbage approach and completely misses the point. Sure, you can completely destroy the session by distilling it into a markdown- but thats like explaining what you did yesterday, vs having an immutable log of what you did yesterday.
This is good, I like it!! I've been mostly working with Claude and Chatgpt for separate stuff, however sometimes one of them is stuck and I could definitely use this to start a session with the other one and see if it figures it out.
Have you tested to see if there was any degradation happening when switching from one model to the other ? I mean some must be inevitable (maybe not!), but how much?
Nars here, one of the co-founders. With longer sessions, the receiving agent sometimes decides to compact it. But the whole transcript still exists and is accessible
In some cases, harnesses try to obfuscate reasoning, but almost everything goes through.