GPT-6 / Claude Opus 5 / Fable 5: Unlimited quotas for top-tier models, web AI directly operates local VS Code projects

Top-Tier Models + More Flexible Usage + Local Project Operations? Don’t Scroll Away Just Yet—Take a Look at My Solution.

I’ve been working on something lately:

Connecting top-tier AI models from the web directly to local VS Code projects.

It’s already up and running, with support for:

  • GPT-6
  • Claude Opus 5
  • Claude Fable 5
  • More advanced models being added continuously

There are two key features:

1. No longer limited by the fixed quotas of traditional plans

One of the biggest problems with using Claude / GPT is:

You’re coding halfway through something, and your quota runs out.

This solution no longer uses the fixed quotas found in traditional official plans.

For developers who use AI frequently, it allows long, continuous conversations and development without constantly checking the remaining quota.

The actual models, available usage, and rules are subject to the website’s real-time status.


2. Web-based AI can directly operate on local projects

I created a VS Code extension that connects web-based AI with the local development environment through MCP.

The AI in the browser can now directly:

  • Read the entire local project
  • Create / modify files
  • Search code
  • Execute commands
  • Run the project
  • Debug
  • Call MCP
  • Use Skills
  • Continue making changes until the task is complete

The entire workflow looks like this:

GPT-6 / Claude Opus 5 / Fable 5
                ↓
             Web AI
                ↓
              MCP
                ↓
             VS Code
                ↓
          Local project

No more copying and pasting code back and forth.
You can simply tell the AI:

“分析这个项目。”

“把这个功能实现掉。”

“运行一下,有错误自己继续修。”

Actual testing
Recently, I used it to build a 「Match-3 automation」 project from scratch.
The AI completed the following on its own:

  1. Read the local project
  2. Modified the Python code
  3. Analyzed the game screen
  4. Wrote automatic recognition logic
  5. Clicked automatically
  6. Performed continuous automatic matches
  7. Continued modifying the code based on the results of each run
    The biggest takeaway is:
    AI is no longer just answering coding questions—it has genuinely started participating in local project development.

You can use models such as GPT-6, Claude Opus 5, and Claude Fable 5.

It’s particularly suitable for:

  • Heavy AI users
  • Programmers
  • Claude Code users
  • MCP / Agent Coding users
  • People who frequently run out of model quota

Current development progress: 98% complete, with support for Skills and MCP tool calls (screenshots of the extension are below)



If you’re also exploring this area, feel free to discuss it with me.

Keep it up, legend! Hope you achieve success in your research soon.

1 Like

:grin:It’s now been successfully developed, with support for Skills and MCP tool calls. I’ve been happily using it for half a month—having unlimited usage is just fantastic.

The goal for the next step is to handle long-running tasks. We’re currently exploring solutions.

Which web-based model has no usage limits?
I feel like 2api is more convenient.

There are many, such as Arena, ChatGPT’s chat mode, Manus, and so on. They can all use the web to connect to local projects for development. I’ve been using them for more than half a month now.

I seem to remember that ChatGPT’s chat also has usage limits; apparently, only the web versions of those Chinese models offer unlimited use.

Doesn’t sub2api cost money? We use so much every day that with a subscription, the daily cost alone would be several hundred. The web AI is completely free, so go wild.

Could you share it, please?

No, that’s not the case. Try it—chat mode can be used continuously. If your account is Plus, you can use the Sol model in chat mode, and it won’t consume your Codex quota (as long as you use chat mode rather than work mode).

sub2api is a project. I’m talking about a web reverse-proxy API—the kind used to connect directly to local agents—which is relatively likely to get your account banned.

arena supports MCP?!:flushed_face:

Haha, of course there are tricks to it. It was fully compatible with arena, and last week you could still endlessly pull the fable model. But this week they fixed the issue that hid the models, so now you can only draw pelicans.

How could there be none???

There should be one. Check your settings or something—there must be some restriction, otherwise 2api would be everywhere.

Arena can’t be reverse-proxied. If it could, I would have set up a reverse proxy to sub2api already. So I developed this MCP specifically to bridge web pages and local projects.

No, that wasn’t the case. I tested it myself for a week with frequent use. Maybe I didn’t use much of the quota or something. In any case, it kept working, and I never got any message saying the quota was full.

Of course—you’re welcome to share your thoughts and discuss them together.

2 Likes

This is absolutely ridiculous. I don’t even know what model you’re using—I suspect you’re using the web version’s Luna model, which would explain how you can use the web version almost unlimitedly with a Plus subscription.

Click the Retry button below and see which model comes up.

By the way, if anyone knows of a solution for long-running tasks, we can discuss it together.