Headroom: a live gauge of your AI chat's remaining context, so you know before the AI starts forgetting. Free & open source. Built after DeepSeek started contradicting decisions settled 20 rounds earlier — no platform shows remaining context. Tokens from a 6-writing-system model calibrated against real tokenizers. Text is read, counted, discarded — no server, no API keys. ChatGPT · Gemini · DeepSeek · Kimi · Qwen · Qianwen · Doubao · github.com/ZM-BAD/headroom

Hey PH! I built Headroom after a long DeepSeek session started contradicting decisions we'd settled 20 rounds earlier — the context window had silently overflowed and nobody had told me. None of the major AI chat platforms show remaining context in the UI, so I made a gauge. A few things I'd love feedback on: 1. It estimates tokens with a linear model over 6 writing systems, calibrated against each platform's actual tokenizer (tiktoken o200k_base, DeepSeek's BPE, Qwen's BBPE...). ±15% on natural language, with known error modes on code-heavy and mixed-script text. Is there a better lightweight approach that fits in a ~70 KB extension? WASM tokenizers were too heavy. 2. The MV3 sidePanel + action.disable() bug — Chrome officially says "Works As Intended", so the workaround is a 3-D per-tab ACL (icon swap + onClicked + setOptions). Any sidePanel devs hit the same wall? 3. Which AI platform should I support next? It's free, no signup, no API key — install from the Chrome Web Store / Edge / Firefox (links in the README). Happy to answer questions all day.

The fact that it runs locally without API keys and never stores your text is a really thoughtful design choice, especially for something meant to monitor sensitive chat sessions.

Headroom

A live gauge of your AI chat's remaining context

A live gauge of your AI chat's remaining context