Token Budget Visualizer

8,192 tokens. Everything must fit. Something has to give.

When I built Atlas, an AI that runs entirely on a phone, the context window stopped being an abstraction and became a housing crisis. The system prompt needs a room. The retrieved documents need a room. The conversation history wants the whole floor. And your new message just walked in the door. Below is the exact negotiation, live.

Type or paste something into the box. Your text claims space using the same heuristic Atlas uses for packing decisions: roughly one token per four UTF-8 bytes, deliberately over-estimated (better to under-fill than overflow). Then drag the history slider up and watch the eviction policy work: when the window overruns, the oldest conversation turns get pushed out first, one by one, while recent exchanges survive. That newest-first packing is smarter than a fixed "keep last 10 messages" because it adapts to how long the messages actually are.

In Atlas this happens before every single model call, in microseconds, with a test suite checking the estimator never drifts from the real tokenizer. The full story: Fitting a Second Brain Into 8,192 Tokens →