An AI coding agent that answers to your machine, not a vendor

What it does

Zero is a Go-based AI coding agent that lives in your terminal. It reads repositories, edits files, runs shell commands, and calls local or remote models through a TUI or a headless zero exec mode. Everything stays local: sessions persist on disk, context never leaves your machine as telemetry, and you pick the provider and the permission level.

The interesting bit

Most agents treat your codebase like a sandbox they own; Zero treats side effects as a privilege to be granted. File writes are confined to the workspace, shell commands and network access are permission-gated, secrets are redacted from logs, and unsafe modes require an explicit opt-in. It is essentially paranoid by default.

Key highlights

• Supports 25+ model providers including OpenAI, Anthropic, Gemini, Ollama, and OpenRouter, switchable mid-session
• Dual interface: an interactive TUI with slash commands and themes, plus a scriptable zero exec with JSON streams and meaningful exit codes for CI
• Extensible through MCP servers, skills, plugins, hooks, and specialist subagents
• Isolated git worktrees and durable local sessions that can be resumed, forked, or rewound
• Native sandbox helpers on Linux and explicit directory grants (--add-dir) to limit write scope

Caveats

• The npm installer is just a wrapper that fetches a native binary; Bun users must manually trust the package to avoid a missing-binary error
• Source builds require Go 1.25+, and Linux users who want native sandboxing must compile a separate helper binary

Verdict

Worth a look if you want a local, model-agnostic agent that treats safety as a first-class feature. Less compelling if you need a graphical IDE integration or a fully vendor-managed service.