# Show HN: Headroom – warns you before your AI chat starts forgetting

> Source: <https://github.com/ZM-BAD/headroom>
> Published: 2026-08-03 13:39:24+00:00

Know when your AI is about to forget.

**English** | [简体中文](/ZM-BAD/headroom/blob/main/README_zh.md)

**Headroom** is a browser extension that gives you a real-time, visual gauge of your context window usage in AI chat conversations — so you know before the AI starts forgetting.

**Zero setup**— No account, no API key, no configuration. Install and it just works.** Real-time gauge**— Watch your context window fill up as you chat. 🟢 Green (plenty of room) · 🟠 Yellow (filling up) · 🔴 Red (time for a new conversation).**Per-round breakdown**— See estimated token counts for every Q&A pair, input and output separately.** Privacy by design**— Conversation text is read transiently to count tokens, then discarded. Only the token counts are stored, locally and optionally in your own cloud storage. No Headroom server exists.

**Supported Platforms & Default Context:** **DeepSeek (1M), ChatGPT (128K), Gemini (1M), Kimi (256K), Qwen (1M), 通义千问 (1M), 豆包/Doubao (256K).**

Context limits are auto-detected per platform and user-overridable. All seven platforms' request parsing and DOM selectors were captured live (2026-06). Built with a platform-agnostic adapter architecture — adding a new AI chat platform is a matter of writing one adapter.

When you're deep in a long conversation with an AI, you've probably experienced this: constraints you set 10 rounds ago get forgotten, earlier decisions get contradicted, and output quality degrades silently. **Context windows have limits** — when they fill up, the AI quietly drops earlier information. None of the major AI chat platforms show you how much room you have left.

**Who is this for?** Professionals who use AI chat as part of their daily workflow — developers, researchers, writers, analysts. If your conversations regularly exceed 20+ rounds for architecture discussions, code reviews, research deep-dives, or technical writing, Headroom helps you maintain conversation quality.

Not a token cost calculator.AI models are getting cheaper by the month — the concern iscontext quality: making sure your AI hasn't silently forgotten something critical.

Install from the

[store badges at the top ↑]— each badge links directly to that browser's extension store.

Headroom is **Manifest V3 only** (MV2 is not supported) and requires a recent browser version: **Google Chrome ≥ 149, Microsoft Edge ≥ 149, Firefox ≥ 151.**

**Known Platform Differences**

| Behavior | Chrome | Edge | Firefox |
|---|---|---|---|
| Icon grays on non-AI pages | ✅ | ✅ | ✅ |
| Panel auto-closes when leaving AI page | ✅ | ✅ | ❌ (content switches to hint page) |
| Panel auto-restores when returning to AI page | ✅ | ❌ (manual click needed) | ✅ (content restores) |

**Edge limitation:** `sidePanel.setOptions({ enabled: true })`

does not auto-restore the panel after switching back to an AI platform tab. Microsoft confirmed this is "by design" ([issue #222](https://github.com/microsoft/MicrosoftEdge-Extensions/issues/222), open since Nov 2024, no fix). Workaround: click the toolbar icon to reopen.

**Firefox limitation:** `sidebarAction.close()`

requires a user gesture and cannot be called on tab switch (Firefox platform limit). The sidebar must be manually closed by the user. When switching to a non-AI page, the sidebar content switches to a hint page via `sidebarAction.setPanel()`

.

Headroom is fully functional with zero cloud setup. For most users, most of the time, that's everything.

Optionally, you can connect your own [Upstash](https://upstash.com/) Redis KV — create a free account, provision a KV instance, and paste the REST credentials into the extension settings (BYOK: the data lives in your own private storage, and the credentials never leave your browser). What connecting adds:

**Settings sync across devices**— thresholds, language, per-platform context limits, and token-coefficient overrides follow you.** Count continuity**— when a platform's history API can't return the full history (pagination caps, interrupted fetches), the cloud record preserves previously counted rounds so your cumulative total never silently shrinks.

Upstash's free tier ([pricing](https://upstash.com/pricing/redis/)) includes 256 MB storage and 500,000 commands/month — far beyond what a single user generates. Headroom stores only token counts per round (no conversation text), so storage is a non-issue (~4 KB per 50-round conversation).

— Next-gen web extension framework (Manifest V3)[WXT](https://wxt.dev/)**Native Side Panel API**— Browser-native sidebar (Chrome`sidePanel`

, Firefox`sidebarAction`

)**Heuristic token estimation**— 6-way per-script coefficient engine (CJK, Kana, Hangul, Cyrillic, Arabic, Latin); character-based for CJK/kana/Hangul, word-based for the rest. No heavy tokenizer bundled (keeps the extension light and model-agnostic)**Upstash Redis KV**— Optional user-owned cloud storage (BYOK model)

We welcome bug reports, feature requests, and code contributions!

**Found a bug?**→[Report an issue](https://github.com/ZM-BAD/headroom/issues/new/choose)** Want to contribute?**→ Read our[Contributing Guide](/ZM-BAD/headroom/blob/main/CONTRIBUTING.md)** Testing help needed?**→ See[open issues](https://github.com/ZM-BAD/headroom/issues)labeled`needs-test`

**Development**

```
npm install
npm run dev            # Dev mode with HMR (Chrome default)
npm run dev:firefox    # Dev mode for Firefox
npm run build          # Production build
```

See [AGENTS.md](/ZM-BAD/headroom/blob/main/AGENTS.md) for full development guidelines and architecture details.

Headroom reads your conversation text only to count tokens — it stores **only
the counts**, never the text itself, locally and in **your own** Upstash KV (if
configured). There is no Headroom-operated server and no third-party tracking.
Full details: [PRIVACY.md](/ZM-BAD/headroom/blob/main/PRIVACY.md).

Copyright 2026 ZM-BAD.

Licensed under the **Apache License, Version 2.0** — see [LICENSE](/ZM-BAD/headroom/blob/main/LICENSE).
You may obtain a copy of the License at

```
http://www.apache.org/licenses/LICENSE-2.0
```

Unless required by applicable law or agreed to in writing, software distributed under the License is distributed on an "AS IS" BASIS.