Platforms selling compute have no incentive to help you use less of it.

In Q1 2026, Anthropic shipped 1M context windows at standard pricing. Dev Twitter exhaled. The context window problem was supposed to be solved.

It wasn't. Chroma's 2025 attention-curve research shows that every frontier LLM decays past roughly 50% utilization with 20 or more turns. That number did not change when the window grew. What changed: the decay moved from visible to invisible. Your conversations are now large enough that you stop noticing the rot until it has already cost you.

Saar started as a token tracker. That was sherlocked the day 1M context shipped. Five Chrome extensions now compete on raw token counting. None of them tell you to stop.

The misaligned incentives problem

Anthropic, OpenAI, and Google are compute businesses. They win when you send more tokens, start more conversations, and pay more per month. They have no product reason to build a tool that tells you to stop. That would be like a bar installing a breathalyzer at the door.

Saar is not a compute business. We win when your AI sessions are sharp, short, and productive. That misalignment is the moat. Providers can ship all the memory features and summarization they want. They still can't tell you with a straight face to start a new chat.

What we built

A Chrome extension that intercepts your Claude conversations locally, counts tokens with a BPE tokenizer, and coaches you in real time. No server, no account, no data leaving your browser. When your context passes 70%, Saar says so. At 90%, it tells you the conversation is spent.

That's Wave 1. Wave 2 is an MCP server on the Anthropic registry and a Tauri menu-bar app that watches Claude Desktop, Cursor, and Claude Code simultaneously. The distribution channel for Wave 2 is governed by the Agentic AI Foundation, a Linux Foundation project co-founded by Anthropic, OpenAI, and Block. Even if Anthropic wanted to sherlock us, the registry they'd have to kick us off is one they share with their competitors.

Who this is for

Solo builders and indie hackers who use AI daily, hit context limits daily, and are tired of figuring out why a 40-turn conversation is giving worse answers than a 5-turn one. People who notice things.

If you think 1M context windows mean you never need to start a new chat, Saar is not for you. That's fine. We'd rather have 1,000 users who care than 100,000 who installed us and forgot.