Local terminal
oap agent chat <class> drives an agent through a full-screen terminal UI — a live transcript,
inline tool sessions, and the same approval prompts you'd see in Slack — with nothing to provision. It's the
fastest way to work with an agent from your own machine.
Interactive vs. scripted
There are two ways to run a session from the CLI, and they're deliberately separate:
oap agent chat <class>— the interactive terminal peer of a Slack thread. It stands up an ephemeral local channel, streams the conversation, and prompts you inline when the agent needs an approval. It requires a real terminal (it refuses a non-interactive one, and points you torun).oap agent run <class> --prompt "…"— a scripted one-shot: apply a session with a fixed prompt and stream its turns, for automation and CI. Flags cap it —--budget-tokens,--budget-turns,--budget-duration— and--no-tailapplies without streaming.
The same rules, a different surface
The terminal is just another channel kind. The agent's tools, identity, authorization, and approvals are identical to what it has in Slack or the browser — only the surface differs. An approval you'd click in a Slack card, you answer at the prompt; a plan the agent proposes renders in the transcript. Nothing about the terminal relaxes a gate.