Quickstart
From clone to chat in minutes. All configuration is arrow-key menus (↑↓ Enter); with no interactive TTY, the CLI falls back to an equivalent numbered menu. You neve…
- Install
git is the only prerequisite. In alpha, review the script on GitHub first if you prefer.
curl -fsSL https://okamiagent.com/install.sh | bash git clone https://github.com/OkamiOps/Okami-Agent && cd Okami-Agent && ./scripts/install.sh - Configure (okami setup)
A wizard walks you through: provider → login → memory → identity → channel. It detects the providers you already have (e.g. a local LMStudio) and prepares okami.yaml.
okami setup - Chat (okami chat)
Opens the full-screen TUI. The session persists — reopen it and the conversation continues where it left off.
okami chat - Run a verifiable task
okami task runs the harness until the exitCriteria are met. Here the task only completes if the file actually exists.
okami task "create hello.txt with 'hi'" -e file_exists:hello.txt - Connect a channel
Set the bot token on the agent and boot the gateway — now you chat over Telegram, with voice and per-chat go/no-go.
okami gateway
What just happened
setup wrote your config; chat opened a persistent session in the default agent's workspace; task ran the harness (not a raw call) and only declared success after verifying the exitCriteria. That's the heart of Okami: completion is a verified assertion, not the model's word.
Inside okami chat
The TUI is concurrent (Claude Code style): you type while the agent works, messages queue, Ctrl-C cancels the turn (doesn't exit) and Ctrl-D exits. Slash commands come from a declarative registry:
/new · /stopnew conversation (archives the current) · cancel the task/model · /thinkswitch model · reasoning effort (low…high)/feedback · /personashape VOICE/PERSONA · session-only overlay/like · /dislike · /differentteach design taste/status · /usage · /toolsstate · tokens and cost · tools/yolo · /undo · /helpauto-approve · revert identity · help