Edgee
One gateway for cheaper, faster, unstoppable coding agents
Start new thread

Edgee Claude Code Compressor - Extend Claude Pro's limit by 26.2%

by•
You're mid-task. Claude is in flow. Then the plan limit hits and everything stops. You know the feeling — the session cuts out, the context is gone, and you're starting over. For heavy Claude Code users, this isn't an occasional annoyance. It's a regular ceiling on what you can get done in a day. We built Edgee's Claude Code Compressor to push that ceiling back.

Add a comment

Replies

Best
Maker
📌

❤️ Today, we're launching the Claude Code Compressor.
I want to show you what it does with a real-world test scenario, so I recorded this video.

I created two separate Claude Code sessions, each connected to a dedicated plan. Same codebase, same task, same instructions: one side standard Claude Code; the other routed through Edgee with compression enabled.

Left side stops at 21 instructions. Right side reaches 26.5.

+26.5% more session before hitting your plan limit.

Here's how it works: Edgee sits between Claude Code and the Anthropic API. Before each request is sent, it strips redundant context, deduplicates instructions, and sends a leaner prompt. Claude sees less noise. You get more range.

To install: curl -fsSL | bash

Then: edgee launch claude

That's it. Free. Takes 30 seconds to set up.

If you're a Claude Code user who's hit the plan wall mid-task, this is for you. If you're running Claude on Anthropic's API and watching your token bill grow, this is also for you.

We've been in beta for a few weeks. Today it's out for everyone.

Hunter

neat product - keep up the great work,  and team 👏👏

Maker

You rock  ! Thank you very much for supporting and highlighting this incredible feature.

does code quality declines??
Maker

 Absolutely not, on the contrary. Compression for Claude Code applies to tool results that are very often too verbose. For example, when the model asks your Claude Code to execute a git log, the model doesn't need unnecessary details. Our compressor cleans up all the polluting elements.

Congrats. A very clever solution to a black-box problem. I’d be interested in learning more about your business model. Will your service offer a paid plan? That would mitigate the impact of the AI provider’s pricing. Or perhaps you monetize the data, since you act as a middleman, which would make it harder for me to choose a solution like this..
Maker
we do not monetize the data, because we do not store the prompts ! Never, ever ! We offer other services for enterprises, like a compressor for agentic use cases, multi LLM, edge tools, caching…
thanks for your answer 👏👏👏

More tokens, fewer plan interruptions 🙌

Maker

 Thanks a lot. Don't hesitate to try it, it's 100% free

Would be great to see a breakdown or visualization of what’s being removed vs kept. That could help build trust in the compression layer.

Maker
you’re right. When you finish a Claude session with Edgee, you can access to a dashboard that shows the savings. And if you activate the debug mode, you also have access to the detail of what we optimized.

Using Edgee already, really great product.

Super simple idea but actually makes a difference on costs

Maker
We are very proud to have your support, especially coming from an entrepreneur like you who is achieving great things.

Super useful in this day and age :) Thanks Sacha and team !

Maker
You're right. There's no need to waste token; a little frugality is welcome these days.

Hit that Claude limit mid-flow way too many times 😅 this kind of compression feels like a simple fix that actually saves real time + money.

Maker

 Simple and efficient. Just a simple CLI install (with brew or culr), then `edgee launch claude`... and that's it, you save up to 50% of token cost :)

Interesting that the fallback sends the original prompt when BERT score is too low. Smart safety net. One thing I'd watch though: Claude Code already runs its own context compression internally, and there are known issues where that causes it to lose track of instructions. Adding another compression layer on top might amplify that. Have you tested how the two interact?