Compresr (YC W26)

We realized that Claude Code and OpenClaw struggle to keep their context concise and prevent context bloat, so we stepped in and built Context Gateway to fix that! Now launched at Product Hunt: https://lnkd.in/gXbxJ_3T. The agents are only as good as their harness - the set of tools they can use to get things done. However, each tool call and each subagent spawned pollutes the agent's context window with thousands of irrelevant tokens. This results in higher token spend, higher latency... And lower accuracy as well! Even the most advanced models with giant context windows suffer from this. GPT-5.4 launch notes show that quality starts decreasing already at 1/30 of the context window capacity. Our proxy reduces context bloat by compressing each tool output before it hits the model, making sure only high-signal info is preserved. Compression is powered by the models from Compresr (YC W26) trained specifically for this task. Our team has completely shifted from vanilla Claude Code to Claude Code through Context Gateway - that's because we have no reason not to: with the proxy, Claude Code is faster, cheaper, and more accurate :)

We keep hearing this in customer calls: “Can compresr be question-specific?” YES! We’re working on it right now (DM me if you want beta access). What does question-specific mean? With the same context, Compresr can compress differently depending on the question: Question-agnostic (cmprsr-v1): compress once → reuse across many queries Question-specific: compress for a specific query → keep only what matters Example in the chart 👇

Bryan Johnson came to Y Combinator today. He and YC often share the same piece of advice — for individuals and for startups, respectively — which is: “Don’t die!” I have something to add to that: don’t let your signal die in messy context — consider context compression :)

At Compresr (YC W26), we build state-of-the-art context compression models — and staying on top of new research is non-negotiable. But startup life = less time to track fresh arXiv announcements. So we built a bot that monitors new arXiv papers and sends us a crisp digest. It’s been so useful that we’re opening it up. If you want to track new papers on prompt compression (soft + token-level) and prompt optimization: Comment “daily” for a daily digest, or “weekly” for a weekly one 🙂 Book a meeting: https://lnkd.in/e_uTGzCX

Compresr (YC W26) is on a mission to make every token count, and we need YOUR input to make it happen! Humanity has produced trillions of tokens of text, yet some documents are used far more often than others. Financial analysts need annual reports from public companies. Lawyers pull up prior cases to prepare for the new ones. High school teachers use a small set of textbooks to plan their lessons. In the era of LLMs, that means the same texts are fed into models over and over again. But what if we could compress each of these texts by 50% or even 70%, keeping them crystal clear for the models? Compress Microsoft’s latest SEC filing just once – so all analysts can include the compressed version in the context window. This would mean reduced compute. Debloated context. Better accuracy. We believe in a future with pre-compressed, LLM-ready versions of the most-used texts on the internet, but as the internet is a bit too large, here is OUR ASK: help us figure out what data to prioritise. If you want to declutter your LLM context and help us stop the token waste globally, please share your use-case here: https://lnkd.in/eYchB7xY Book a meeting with one of the founders: https://lnkd.in/e_uTGzCX