If you're hitting Claude Code's context limit, check your terminal output before blaming your project size. Every command Claude runs dumps raw output into your context window. git status : 76% of it is noise. A failed cargo test : 155 lines becomes 3. A grep across your codebase... : same story You're not running out of context because your code is too big. You're running out because the commands are bleeding you dry before you get anywhere. RTK intercepts that output, compresses it, and hands Claude a clean version instead. 60–90% fewer tokens. Same information. Runs in the background after a one-command setup. Real numbers from 2,900+ commands: average 89% compression across the board. MIT license. Free. 39k GitHub stars. rtk init -g That's it. Restart Claude Code and every command is compressed automatically.show more

Repo:

this is actually helpful

Appreciate it man!

the real issue isn't even hitting the limit, it's that terminal noise destroys the model's attention. 1M context on opus 4.7 is useless if 80% of it is raw git logs.

rtk init -g and done?? no way

@charliejhills yep, been there. nothing like getting buried in logs just to find that one tiny error. makes me wish for an automatic "important stuff only" filter tbh

Amazing share

Thanks man

Claude suffocating on logs while I’m confused 😭

It’s a little confusing what extra steps windows users need to do?

rtk sounds like a context band-aid, tho. what's next for terminal output compression? 💀

This is the part most teams miss. Context limits usually look like a model problem, but they are often an observability problem. Once command output gets compressed into the useful bits, agents stay effective a lot longer without changing the underlying workflow.

terminal output is the silent context killer. my rule now: commands should summarize by default, dump artifacts to files, and only paste the 20 lines that matter back into the agent. otherwise every test run becomes accidental prompt injection against your own context window.

The terminal output isn't the real issue. Most context bloat comes from Claude re-reading the same files every turn when it loses track of what changed.

The real issue isn't output verbosity, it's that most devs don't realize context is a resource to manage like memory. What's your strategy for pruning it mid-session, or do you just let it blow up and restart?

most people think the problem's with claudes capabilities, but a lot of this 'noise' is just cruft we bring to the table ourselves.

wait 155 lines → 3 is actually insane !!!

Terminal output is the silent context killer most people blame on project size. The bet with RTK is staying lossless on error paths, where the rare line explaining why a test failed has 100x the value of routine 'all good' chatter.

We’re moving from "Chatting with AI" to "Giving AI the terminal." If you aren't optimizing the feedback loop with tools like RTK, you’re essentially running your LLM on a dial-up connection. This is how you scale complex repo navigation without hitting the wall.

We rolled out cost management in AgentsRoom a while back 💸 You can track Claude Code token usage, keep an eye on spend, and stay in control as your agents run. ➡️

well share bro

Context blowups are usually an output hygiene problem before they are a model problem. Compressing command chatter at the source is a smart fix, especially for long agent loops where noisy tests and grep output quietly eat the whole budget.

Yep. Compression fixes the garbage going into context, but I still want an odometer for the session itself. Terminal noise is exactly the kind of token leak that should be visible live: