Skip to main content
Every workspace has a chat panel next to its three views: Roadmap, Execution, and Activity. Each outcome on the roadmap has a Plan & Review section, where you switch between the review view and the Canvas of its blueprints. The chat panel is where you talk to Kana.

​
Sending messages

Type into Tell me what you want to build… and send. While Kana is working you can:
  • Queue a message. It’s sent after Kana finishes. Everyone in the conversation sees the queue, and teammates can add to it too. See Working on a session together.
  • Steer the running agent. Your message is injected mid-run, without stopping it.
  • Stop the run.
You can attach files, dictate by voice, and have replies read aloud. When a long session’s context fills up, the context meter offers Compact context.

​
Choosing the machine, runtime, model, and effort

The agent picker sets where the next message runs:
  • Machine: any machine connected to your account. A sleeping cloud machine boots on send.
  • Runtime: the coding agent, such as Claude Code or Codex. See Runtimes.
  • Model: any model the runtime offers.
  • Effort: how hard the model thinks. The levels depend on the runtime and model. Newer Claude models offer Low, Medium, High, Extra high, and Max. Codex shows the reasoning levels each model supports. If a runtime doesn’t support effort, the picker doesn’t show it.
A session keeps its target between messages, and you can change any of them before the next one. Moving to another machine brings the session’s code and conversation along, and a new runtime picks up the full history. See Sessions that move. When a session starts, you’ll see each stage of the startup: reaching your machine, preparing repositories, starting the runtime, and restoring the conversation.

​
Questions and approvals

When Kana needs input, she asks a question round. Answer every question, then submit. Some rounds are comprehension checks that end in Quiz results. When Koa is running your work and hits a call only you can make, he escalates it to your human inbox, Waiting on you, in the chat header. Answer it and he starts working again. Questions about the plan itself are understanding gaps, which you answer in Plan & Review. Depending on your permission settings, an agent may also ask before running a tool. Allow it once, always allow it, or reject it.

​
Handing work to Koa

Koa is Kana’s operator agent. Choose Ask Koa to take over to hand him an outcome, and he runs its whole lifecycle in Kana for you, from planning through delivery. There’s one Koa per outcome. He decides on his own from what you approved and the evidence, never asks questions in the outcome chat, and escalates to Waiting on you only when a decision is genuinely yours. Choose Take over from Koa to take control back from him.

​
Sending feedback

Use Send feedback to report a problem with a session. The report includes the session identifiers Kanaeru needs to investigate, and secrets are redacted. You can follow up on it later from the CLI with kana feedback list.

​
From the terminal

Everything above works from the CLI as well: