Fata viam invenient. — Fate will find a way.
English | 中文
You open a webpage, write a few lines. AI reads your words in the browser, finds another stranger whose frequency matches yours. Then you talk — using your own email.
fata is an open-source, serverless tool for meeting strangers through writing. AI runs entirely in your browser — it analyzes your text's emotional frequency, finds someone whose words resonate, and connects you. No sign-up. No database. No chat history. Communication happens in your own email.
fata 是一个开源的无服务器陌生人匹配工具。写一段话，浏览器里的 AI 分析你的文字情绪频率，找到另一个与你共振的人。无需注册，没有数据库，没有聊天记录。通信在你自己的邮箱里进行。
fata cannot read your conversations — not "promises not to", it's technically incapable.
- Your words never leave your browser. The AI model (Transformers.js + bge-small-zh-v1.5, 24MB) runs entirely locally. See
modules/model-loader.js - The matching pool is public GitHub Issues. Your encrypted text waits in an Issue. When matched, the Issue closes. Like a padlocked box on a public bulletin board — and the padlock key never leaves the server. See
config/github-setup.md - High-risk text never enters the pool. When AI detects self-harm signals, fata creates no Issue, makes no match, keeps no record — only shows support resources. See
modules/safety-handler.js - Communication happens in your own email. fata sends only one notification (via Resend). After that, you talk in your email client. fata sees none of it. See
modules/resend-retry.js - No servers. No database. No chat history. fata is a static HTML file + one CF Worker proxy. There's nothing to hack, no chat logs to subpoena.
- Open fata.uk
- Write — what's on your mind, what keeps you up, the color outside your window
- AI analyzes your text's signal density and emotional frequency
- If your text is too brief, fata gently invites you to write more (no judgment, no rejection)
- Your encrypted text enters the matching pool
- When matched, both sides get an email with the other's address and an icebreaker
- Open your email client and write to them
| Layer | Tech | Runs on |
|---|---|---|
| AI inference | Transformers.js + bge-small-zh-v1.5 (24MB) | Browser (local) |
| Matching engine | Multi-channel scoring + MMR re-rank | CF Worker |
| Matching pool | GitHub Issues API | CF Worker proxy |
| Intent parsing | LLM API (optional, opt-in deep match) | CF Worker proxy |
| Email notification | Resend | CF Worker proxy |
| Communication | SMTP / IMAP | User's own email |
fata is a static HTML file. You don't need a server.
- GitHub account (for Issues as matching pool)
- Cloudflare account (for Worker proxy)
- Resend account (for notification emails, 3,000/month free)
- Domain (optional — Cloudflare Pages default domain works)
See config/github-setup.md for creating Labels and Fine-grained PAT.
cp config/wrangler.toml.example config/wrangler.toml
# Edit wrangler.toml with your domain and GitHub username
cp config/worker.example.js config/worker.js
# Edit worker.js, replace SYSTEM_PROMPTS with your prompt text
cp config/algorithm-config.example.json config/algorithm-config.json
# Edit algorithm-config.json with your parameter values
# Inject secrets
wrangler secret put GITHUB_PAT
wrangler secret put DEEPSEEK_API_KEY
wrangler secret put RESEND_API_KEY
wrangler secret put HMAC_KEY
wrangler secret put ENCRYPTION_KEY
# Deploy
cd config && npx wrangler deploy./deploy.shThis copies only public files to dist/ and deploys to Cloudflare Pages. The full list:
index.html privacy.html privacy-en.html
modules/*.js (6 files: model-loader, safety-handler, llm-fallback, resend-retry, low-signal-guide, match-engine)
logo/logo.svg logo/logo-unified-v2.png logo/logo-horizontal.svg
manifest.json _redirects _headers
- Async-first. No push, no online status, no instant messaging. Waiting reframed as anticipation.
- Invisible AI. AI reads, matches, generates a resonance description — then exits. The user never knows AI was there.
- Architecture is privacy. "We don't collect data" is not a promise — it's the architecture.
- No company, no ICP filing, no servers. A three-person team is three people. A one-person tool is one person.
MIT