Pocket：浏览器直连的 BYOK AI 聊天应用，无后端、本地存储记忆

Live: pocket-tau-sepia.vercel.app
New to Pocket? Try the five-minute first-use review — no API key required.
Already using it? Report a reproducible provider, browser, or UX problem. Please omit API keys and private chat content.

Browser-only AI chat. Paste a supported provider key — Anthropic, OpenRouter, Groq, Together, or a compatible endpoint — and the model picker fills itself. Chats, attachments, and remembered facts are stored locally in IndexedDB. When you use a hosted model, Pocket sends requests directly from your browser to the selected provider; Pocket itself has no backend.

Why I built this

Buying API access shouldn't mean handing your data to a third-party chat wrapper or paying $20/month for a UI someone else owns. I wanted a clean place to talk to whatever model I was paying for, with my own memory, no backend, no subscription.

So:

• One key, any model. Paste, auto-detect, chat.
• Yours alone. No server. Browser → provider, that's the only network hop. Clear browser data and Pocket forgets you.
• Memory that sticks. The assistant ends turns with MEMORIZE: … lines. Pocket harvests them silently into a local fact store, recalls top-matching ones on every new turn.
• Token-aware. Sliding window with summary compaction. Optional heuristic prompt compression. Anthropic prompt caching on by default.

Works as a standalone chat app. Also designed to pair with zhub — paste a zk_ URL+key and Pocket talks to your own AI living on a $5 VPS instead of a hosted provider. Same surface either way.

What it actually feels like

A clean chat. Direct, no preamble, no "sure, here's…" filler. Memory carries between sessions on the same device. Switching models mid-conversation keeps the thread.

Provider matrix (browser-direct)

Paste a key, the provider is auto-detected from the prefix, the model picker fills with that provider's catalog. Models from different providers can be switched mid-conversation; memory and threads survive the switch.

Memory layer

The default system prompt instructs the model to end any turn worth remembering with MEMORIZE: <fact> lines. Pocket:

1. Strips those lines from the user-visible reply.
2. Stores each fact with simple keyword tags.
3. On the next user message, recalls the top 8 facts whose tags overlap and prepends them to the system prompt.

You can browse, search, and delete facts in Settings → Memory. They never leave your device.

Cross-session recall is a separate, opt-in layer: when you message any model, Pocket scans every chat in this browser for relevant snippets and includes them as context. So whichever provider/model you switch to, it walks in knowing what you've talked about before.

Token strategy

Three layers, top to bottom:

1. Sliding window — keep the last N turns full (default 16); compact older turns into a single summary block.
2. System-prompt cache (Anthropic) — cache_control: ephemeral cuts cost on repeated context by ~90%.
3. Heuristic compression (off by default) — strips filler words and applies abbreviations to user prompts. Lossy, opt-in.

Live token estimate is shown under each assistant reply.

Personality / "Presence"

Pocket carries itself a certain way across every model you plug in — direct, no preamble, no closers, willing to say "I don't know," matches your cadence (English / Hinglish / whatever). That posture lives in a single editable system prompt, sent first on every turn before recalled memory.

Edit it to taste. Reset to default any time.

Privacy

• API keys stored in IndexedDB. Optional encryption at rest (PBKDF2 + AES-GCM, prompts for passphrase per browser session).
• No telemetry. The only outbound HTTP from Pocket is to the provider you configured.
• vercel.json ships strict headers (X-Frame-Options: DENY, Referrer-Policy: no-referrer).
• Clearing browser data wipes everything Pocket knows.

What it actually costs you

Whatever the provider charges. Pocket adds nothing.

That's my provider console after a real testing session: my own key, my own bill, no subscription, no SaaS markup. Three cents for six conversations on a flagship reasoning model.

Run locally

npm install
npm run dev    # vite at http://localhost:5173
npm run build  # static bundle in dist/

Deploy

Pocket is a static SPA. Anywhere static works:

• Vercel — vercel (zero config; vercel.json already in repo)
• GitHub Pages — push dist/ to gh-pages branch, or use the included Actions workflow
• Any static host (Cloudflare Pages, Netlify, S3 + CloudFront, your own nginx)

Stack

Vite + React + TypeScript. Tailwind for styling. Dexie (IndexedDB) for local storage. No bundler-side proxy, no Node server, no backend at all — every API call goes browser → provider directly.

License

MIT. See LICENSE.