CLI

Use the terminal when you want to work alongside an agent: inspect a repository, edit files, research a question, or review its actions as they happen.

Start a conversation

Open Jazz from the directory you want the agent to work in:

jazz

On first use, choose Start setup to create an agent and configure its provider. See Quick start if Jazz is not installed yet.

After setup, the home screen offers your recent agents. Select one with the arrow keys, type a request, and press Enter. Tab opens the full agent picker. Type / to find home commands such as /resume for earlier conversations, /start or /new for a fresh conversation, and /create to create an agent, and /settings for providers, keys, appearance, and limits.

To open a particular agent directly:

jazz agent chat assistant

Replace assistant with the name you chose. Try a request that uses your current checkout:

Review the last five commits. Explain the behavior changes and flag anything risky.

Jazz shows the response and tool activity as the agent works. Ask follow-up questions in the same conversation; the agent keeps the earlier context.

The closing turn summary shows time, model steps, input and output tokens, and cost. When generation timing is available it also shows N tok/s: output tokens per second of model generation, excluding tool work and time before the first token. It ends with the local time the turn finished, such as 07:28 am. The same summary, finish time included, remains available after resume.

Give it files

Type @ to choose a file, or name it directly. Relative paths use the directory where you started Jazz; quote paths containing spaces.

Summarize @README.md and list the setup steps.
Compare @"design notes.md" with the implementation and identify gaps.
Explain the errors in @~/Desktop/build.log and inspect the relevant source files.

Images, PDFs, audio, and video can also be attached this way. Media support depends on your model; companion models let you assign specialists for analysis or generation. Web search and integrations need their own configuration.

Approve actions and steer the work

Chat starts in safe mode: read-only and low-risk actions proceed, while high-risk actions ask for approval. Review the proposed tool and arguments before accepting. Reject a step when it is wrong; the agent receives the refusal and can revise its plan.

/mode yolo allows tools to run without asking, including shell commands. /mode safe restores approvals. Shift+Tab switches modes. See Approvals for the policy and command-classification details.

While a run is active, type and submit a correction, such as:

Keep the public API unchanged. Fix only the failing test.

The agent receives submitted messages at its next tool boundary. To interrupt a run, press Esc twice quickly; the first press shows a hint. An open picker or overlay consumes Esc first. In the fullscreen interface, interrupting with submitted messages waiting starts the next turn with those messages. An interrupted action may already have made changes; inspect them before retrying.

Keep and manage conversations

Conversations are saved locally. Memory can carry preferences into future conversations; it is separate from conversation history.

Work beyond one turn

• Goals and loops: use /goal <objective> to keep working until completion is verified, or /loop 10m <prompt> to check on a schedule.
• Remote handoff: /detach <host> continues the conversation on a registered SSH server.
• Headless: run an agent from scripts with jazz run.
• Scheduled runs: save a repeatable job and put it on a clock.

Use /help for commands and shortcuts, or /help goal for a specific command. The command reference lists every CLI flag and chat command.