GitHub

AI coding agent for your terminal

Built for understanding what you are building.
Anti vibe-coding approach.

  • Open source
  • MIT
  • Runs locally
  • No account

Demo transcript summary: Zeta explores a repository, asks clarifying questions, then implements a small CLI flag with a test. Playback is decorative; install and run zeta locally for the real agent.

§01

modes

Press shifttabto cycle build → plan → ask without leaving the prompt.

build

edits need yes

Clarifies with ask_user when the tree can’t settle a choice. Shell and every file edit need a yes.

ask

no writes

Search and explain only. No edits, no shell. Switch to Build to apply changes.

plan

approve → build

Explores the tree, then emits a plan. Approve picks a build model and starts implementing.

§02

why zeta

Nothing silent

Shell and file writes ask first. Bash can be allowed for the session; every edit still needs a yes.

Your models

All models via BYOK or OAuth - bring your own keys and providers. Configure in /config; switch with /model.

Local by default

Chat history stays under ~/.zeta. Resume anytime, auto-compact when the window fills.

/review

Bundled playbook for a strict code-quality pass on the current branch - structure first, nits later.

§03

reference

input

  • enter

    send message

  • shifttab

    cycle build → plan → ask

  • shiftenterorctrljoraltenter

    insert newline

  • escorctrlc

    cancel turn (or quit when idle)

  • prompt history

  • mouseorpguppgdn

    scroll transcript

commands

  • /

    command autocomplete

  • /config

    manage providers & models

  • /model

    switch model

  • /resume

    open a previous session

  • /compact

    summarize older context

  • /review

    strict review of the current branch

  • /clear

    start a new session