🌶️ Zest CLI from Spicy Lemonade 🍋

Small language model. Runs offline. Runs locally on your Mac. No cloud. No API keys. Privacy first. No tracking.

Zest CLI translates natural language into shell commands using a local small language model. MacOS only. Learn more at zestcli.com.

🚀 Beyond the Basics

While the GIF shows simple file management, Zest is designed for a variety of domains:

• Networking: "Show all active listening TCP ports"
• Docker: "Show the logs for the Docker container my-app"
• Git: "Show the most recent commit that modified the file components/pages/AboutPage.tsx"
• System: "Check my disk usage sorted by the top 5 largest directories in my home folder"

Note: Zest works best with single requests. i.e. show the top 5 XXX is better than show the top 5 XXX and then show the bottom 5 YYY

Why Zest CLI?

• Small language model — a purpose-trained SLM optimised for CLI command translation
• Runs offline — runs entirely on your machine with no network calls at inference time
• Runs locally — uses your Mac's built-in GPU via Metal, with automatic CPU fallback; no dedicated GPU needed
• No cloud — no cloud dependencies at inference time
• No API keys — no accounts, tokens, or subscriptions required
• Privacy first — your commands and queries never leave your device
• No tracking — zero telemetry, user data, or usage data collection

🏗 Model Architecture

Zest CLI runs a single model, Qwen3.5-9B, fine-tuned for natural language to CLI translation with supervised fine-tuning (SFT) on ~74,000 real world examples, then aligned with Direct Preference Optimization (DPO). The model is quantized to Q5_K_M in GGUF format (~6.6GB) and runs on a bundled llama.cpp inference engine, so there's nothing extra to install.

The core agent runs offline on your Mac (privacy-first, no tracking). The training pipeline utilises Google Cloud Storage (GCS) to manage datasets before they are loaded into Google Colab for GPU based fine-tuning.

---

🛠 Tech Stack

• Terraform: State management and infrastructure provisioning across environments.
• Google Cloud Platform (GCP):
  • Cloud Storage (GCS): Persistent storage for training datasets and artifacts.
  • IAM: Role-based access control for engineering and operations teams.
  • Cloud Billing: Cost tracking and budget alerts via email.
• Cloudflare:
  • DNS & Domain Management: Domain zestcli.com, and DNS.
  • Security & Performance: Website passed through cloudflare proxy
• GitHub Pages:
  • Free tier hosting on GitHub Pages.
• Zoho Mail:
  • Email Hosting: Custom-domain email for internal and operational communication.
• Resend:
  • Transactional Email: Reliable delivery for product emails (auth, notifications, receipts).
• Polar.sh:
  • Payments & Checkout: Subscription management and checkout flow for monetization.

New developers must be onboarded with IAM permissions and API keys before working with this infrastructure. See docs/setup.md for setup, deployment, IAM, and secrets configuration.