Stop using ChatGPT for shell commands. Type what you want in plain English, get the shell command instantly.

Live: nl2shell-web.vercel.app

A web interface for the NL2Shell model — a fine-tuned 800M parameter model (Qwen3.5-0.8B) that translates natural language to shell commands. No API keys, no subscription.

Browser → Next.js on Vercel → Gradio Space API → NL2Shell Model (HuggingFace)
                ↓
        Supabase (feedback + analytics, optional)

• Text input with Enter-to-submit
• Voice input via Web Speech API (Chrome/Edge)
• Dangerous command detection (22 patterns)
• Copy-to-clipboard
• Feedback collection (thumbs up/down + corrections)
• Command history (last 20, session-only)
• Structured logging (Vercel-compatible JSON)
• Security headers (CSP, HSTS, X-Frame-Options)
• Rate limiting (30 translations/min, 10 feedback/min per IP)

git clone https://github.com/nl2shell/nl2shell-web.git
cd nl2shell-web
bun install
bun dev

Open http://localhost:3000.

Create .env.local:

# Optional: HuggingFace token (avoids anonymous rate limits)
HF_TOKEN=hf_your_token_here

# Optional: Supabase (enables feedback persistence + analytics)
SUPABASE_URL=https://your-project.supabase.co
SUPABASE_ANON_KEY=your-anon-key

# Optional: IP hashing salt for privacy
IP_SALT=your-random-salt

The app works without any environment variables. Supabase integration is opt-in — without it, feedback is logged to stdout only.

If you want to persist feedback and track usage:

1. Create a Supabase project (free tier: 500MB)
2. Run supabase/schema.sql in the SQL Editor
3. Add SUPABASE_URL and SUPABASE_ANON_KEY to .env.local

Useful queries once data is flowing:

-- Daily translation count
SELECT date_trunc('day', created_at) AS day, count(*)
FROM translations GROUP BY 1 ORDER BY 1 DESC;

-- Feedback summary
SELECT rating, count(*) FROM feedback GROUP BY rating;

-- Corrections for training data
SELECT query, command, correction FROM feedback
WHERE rating = 'down' AND correction IS NOT NULL;

Vercel Hobby plan (free):

• 100K function invocations/month — blocks after that (no surprise bills)
• 100GB bandwidth/month

HuggingFace Space (free CPU tier):

• Sleeps after 48h inactivity, wakes on request (~30s cold start)
• CPU inference: 1-15s per query depending on length
• Queues under load — won't cost money, just gets slow

Rate limiting:

• 30 translations per IP per minute (in-memory, resets on cold start)
• For production at scale: replace with Upstash Redis (free: 10K commands/day)

If you get a million users: The HF Space will queue and timeout before billing becomes an issue. Vercel Hobby blocks at limits. Upgrade path: Vercel Pro ($20/mo) + HF Dedicated Inference ($0.60/hr GPU).

curl -X POST https://nl2shell-web.vercel.app/api/translate \
  -H "Content-Type: application/json" \
  -d '{"query": "find all python files"}'

Response: {"command": "find . -name '*.py'", "meta": "..."}

curl -X POST https://nl2shell-web.vercel.app/api/feedback \
  -H "Content-Type: application/json" \
  -d '{"query": "list files", "command": "ls", "rating": "up"}'

• NL2Shell Model — The fine-tuned model
• NL2Shell Dataset — 12,834 training pairs
• Gradio Demo — Direct model interface
• Vox CLI — Terminal client
• CloudAGI — Agent Credit Economy

MIT