Stop babysitting your terminal
With most tools, if the agent hits a question and you are not watching, the run stalls. You come back to a blinking cursor and nothing done. Claude Code is the same: it is tethered to your terminal, your laptop, and your attention. Miss a prompt, lose the run. Close your laptop lid, lose the run. Skydive agents run in the cloud and reach you wherever you are. If the agent needs a decision, instead of stalling, you have taught it to text you. You reply, it continues. Close the laptop, the work does not stop. Laptop open — start the task:ctrl+t. The agent reads your repo through Portal, digs into the codebase, and starts on the job. You watch the diff stream in the file pane and steer it on one edge case. Then you close the lid and go to standup.
Laptop closed — agent still running:
It does not need your machine anymore. It has the context it pulled and keeps working in its cloud sandbox.
It hits an ambiguous interface — two ways to handle the retry logic, not sure which you want. Instead of stalling, you have taught your agent to text you for decisions like this: “payment retry: service layer or controller?” You reply from your phone: “service layer, keep the controller thin.” It continues.
The rest of the morning:
11am, Slack. Your agent posts in #backend that the PR is drafted. You reply in the thread: “also add rate limiting to the new endpoint.” It picks it up without you leaving Slack.
11:30am, away from your desk. You text it: “where are we?” It replies over SMS: payment module done, rate limiting in, tests passing, PR is open. You text back: “hand it to the PR review agent for a security pass.”
It messages your code review agent — a separate agent with its own persona and its own model, tuned for security and correctness — and hands off the PR. The reviewer picks it up, runs through the diff, and posts its findings in Slack.
12pm. You read the review over lunch, make one call, and merge.
One task. A handful of messages across terminal, Slack, and SMS. The agent asked one question the whole morning and you answered it from your phone between meetings.
What the terminal is good at
The CLI is one of several ways to reach the same agent. It has the same memory and the same conversations as web chat, Slack, and SMS, so you can start a task here and finish it anywhere. Three things it does that the other channels cannot:Your local machine
ctrl+t grants the agent access to this machine over Portal: your repo, your localhost, your tools. Off until you grant it, per agent, revocable the same way.Reviewing diffs in place
ctrl+g opens the file pane: every edit the agent made, as reviewable hunks. Comment on a line and send the comments straight back into the conversation.The agent's own sandbox
/sandbox opens a live terminal inside the agent’s cloud machine, or runs one command there and streams the output.Scripts and pipes
chat -p sends one prompt and prints the reply. No TTY needed, so it fits CI, cron, and anything you can pipe into.Sharing this machine
Portal is how an agent reaches your own computer: your repo, yourlocalhost, the tools you have installed. There are two ways to open it, and they grant the same thing.
Skydive Desktop keeps your machine connected in the background. Approve an agent once and it can reach you whether or not a terminal is open.
This CLI, if you would rather not run a second app. ctrl+t opens Portal for the session and /portal grants the agent you are talking to. skydive chat --share-machine does both at launch, and skydive portal open --agent <name> runs it headless with no TUI at all.
Either way it is default-deny and per agent. Nothing reaches your machine until you say yes, and you can take it back at any time.
Get started
Install and sign in
Install script or npm, the beta channel, and your first
skydive chat.Chat TUI
Keybinds, slash commands, images, the file review pane, local shell.
Scripting
One-shot prompts, JSON output, and pipelines with
chat -p.Command reference
Every command, flag, and environment variable.
Related
Reaching your computer
How local machine access works and how the agent uses it.
Agents overview
What the cloud sandbox, memory, and connections are.
Routines
Agents that run without you in the loop.
Platform CLI
The other CLI: the one the agent runs inside its own sandbox.