zox
Get started

The local coding agent harness

Talk to an LLM that can read, edit, and run tools on your repo — with permissions, sandboxing, sessions, and observability. BYOK.

bun add -g zox-code
zox --model openai/gpt-4.1
zox  ·  anthropic/claude-sonnet-4  ·  worktree

>

What is Zox?

Zox is a coding agent harness. The model does not “have” your repo — the harness holds the session, assembles context, streams the model, runs tools under policy, and records traces.

It is not a hosted chat product. The agent loop lives on your machine.

  • Local-first — Server binds 127.0.0.1; you own the keys
  • Worktree sandbox — Default isolation via git worktree
  • Permissions — Approve writes and shell before they run
  • BYOK providers — Anthropic, OpenAI, Google, Groq, OpenRouter, …
  • MCP + skills — Extend tools and reusable SKILL.md packs
  • TUI + SDK — Ink terminal UI and typed @zox/sdk client
Read docs →

How it works

  1. You type a request
  2. Optional prompt guardrail (Jev) can allow / ask / deny
  3. Model streams; may emit tool calls
  4. Each tool: permission → hooks → sandbox → execute
  5. Results return to the conversation; loop until done
  6. Usage, traces, transcript stored on the session
You (TUI / REPL / SDK)
        │  HTTP + SSE
        ▼
   Zox server  ──► LLM provider
        │
        ├── tools
        ├── sandbox
        ├── hooks
        └── SQLite sessions

Built for privacy first

Zox does not host your code. Keys stay in your environment.

The server runs on your machine. Session data lives in local SQLite.

FAQ