Skip to content

Agent CLIs and modes ​

Splitkick runs the agent CLIs you already have. It never sees a password and has no account of its own.

Agents ​

AgentNote
Claude CodeYour Claude Pro or Max plan through the signed-in CLI. Splitkick only reads the login; it never refreshes it, so you do not get logged out elsewhere. You can use several Claude accounts (Beta).
CodexChatGPT plan through the signed-in CLI.
CursorCursor CLI.
OpenCodeOpenCode CLI.
Pinpm-managed CLI.
Command Codenpm-managed CLI.
Groknpm-managed CLI.
Kimi Codenpm-managed CLI.
AntigravityAntigravity CLI.
CodeRabbitReview only: CodeRabbit CLI runs a review of the worktree.

Settings show, per agent, whether the CLI was found, which account is signed in and how much of your plan is used. For Claude you can add more accounts in Settings → Claude → Accounts; see Several Claude accounts.

TIP

Agent profiles (persona, graph permission, default model) are in Graphs and agents.

The model picker: one tab per agent, then the models of that agent

Modes ​

Every session has a mode. You can change it between turns.

The mode menu in the composer: Plan, Build, Yolo

  • Plan. The agent reads and thinks but does not change files. It ends with a plan that you approve or send back.
  • Build. The agent edits files and runs commands, and asks you before anything it is unsure about. Agent questions appear in the session and on the phone.
  • Yolo. No questions. Use it in a worktree you can throw away.

A plan waits for you at the end of the session. Approve runs it in Build mode, YOLO runs it without questions:

A plan waiting for approval, with the buttons YOLO and Approve

When an agent asks a question, the session waits and shows the choices:

An agent question with single and multiple choice, and the buttons Answer and Skip

Model and effort level are set per session and can differ per mode.

While a session runs ​

  • Fork a session from any message, also while it runs, to try a different direction in a new session. See Fork a session.
  • Cancel safely: your prompt goes back into the composer, the conversation is kept.
  • Queue the next prompt: a prompt sent while the agent works, or just after a reply, waits in the queue and runs when the agent is free.
  • Write the next question on the follow-up list and send it when you are ready.
  • Star, pin, label, pause or sort sessions; a draft marker shows unsent text.
  • Home and Recent list sessions across all projects; search covers names and messages.