kon

A coding agent for your terminal. Starts in 27 ms. One binary. Every OS.

kon reads, edits, and runs commands in your project until the work is done. One 12 MB executable: no runtime, no daemon, no warm-up.

kon.kitsu.red

Install

Linux and macOS:

curl -fsSL https://kon.kitsu.red/install.sh | sh

Windows, in PowerShell:

iwr -useb https://kon.kitsu.red/install.ps1 | iex

With Go:

go install kon.kitsu.red/cmd/kon@latest

Get started

Run kon in any project. Connect a provider with /login, pick a model with /model, and start typing. Type @ to mention a file, and / for commands.

kon has no sandbox and does not ask before running commands, so use it on work you keep in version control. The user guide covers the rest; /help opens it inside kon, and kon reads it to answer questions about itself.

Why kon

• Four sharp tools. read, write, edit, and shell, after pi. Context goes to your code, not a tool menu.
• Instant. 27 ms from launch to first frame. No update check, no catalog refresh, and no session scan on launch.
• Native everywhere. Linux, macOS, and Windows on x64 and ARM64. No WSL, no account, no plugins.
• No lock-in. Sessions are plain JSONL you can read, grep, and keep. OpenAI, Anthropic, OpenRouter, Ollama, or any OpenAI-compatible endpoint.

Scripting

kon run sends one prompt and streams the reply, so kon fits in pipelines:

git diff | kon run --stdin review this

Add --format json for one event per line. See scripting.

Editors

kon acp speaks the Agent Client Protocol on stdin and stdout, so editors such as Zed can run kon as their agent. See editor integration.

Documentation

• User guide: the list of its pages
• Getting started: install and first session
• Working with kon: prompting, sessions, jobs, and subagents
• Configuration: providers, models, and project instructions
• Scripting: kon run and kon md
• Editor integration: kon acp for ACP editors
• Reference: every command, key, tool, and setting
• Development: building and contributing
• Architecture: how the pieces fit together
• Session format: the on-disk JSONL contract
• Storage migrations: version upgrades and recovery
• Rendering performance: how the TUI stays fast

License

MIT