Each task runs in its own git worktree, so your checkout is untouched, and every turn is checkpointed so you can roll back. For anything you are unsure about, use Plan (read-only) or Edits, where anything else the agent wants to run comes to you first.

Can I watch what it is doing?

Yes. A working card shows a live one-line status, the full transcript is one click (or V) away, and T opens the session in your own terminal.

Does it need to stay open?

wisp runs the agents, so it has to be running somewhere. Run it on a desktop or a VM through Machines and your laptop can sleep.

Guide

Run coding agents in the background

Headless coding agents (claude -p, codex exec) can work without you, but they can’t ask a question, nothing tells you when they finish, and what they leave is a long transcript. wisp runs them in the background with a way to ask you mid-run, notifications on your desktop and phone, and a one-screen report when each one is done.

Source on GitHub

Free and open source · Claude Code, Codex, Cursor CLI, opencode, Antigravity CLI, Grok Build · macOS, Linux, Windows (beta) · Node 20+

A wisp report for a finished background task: headline, numbers, the real diff and the test run

What plain headless runs miss

When the agent…	Plain `-p` / `exec`	With wisp
needs a decision	guesses, or stops	asks through an `ask_user` tool; the run pauses and you get a notification
hits a permission prompt	the call fails	Allow / Deny on the card, or Allow & retry
finishes	you check the terminal	desktop and phone notification
reports back	thousands of lines of log	a headline, the real diff, the tests, next steps
edits files	in your checkout	in its own worktree, with a checkpoint per turn

Start one

npx wisp-agents "fix the flaky auth test"            # this folder, the agent you last picked
npx wisp-agents --agent codex --mode plan "audit deps" # read-only
wisp "bump node to 22" --wait --json                   # block until done, for scripts

Or open the app with npx wisp-agents and use the launcher. Pick a permission mode per task: Plan (read-only), Edits (file edits allowed, the rest asks you) or Full auto.

While you’re away

Phone. Questions and finished tasks go to ntfy or Telegram. Tap an option to answer, or start a task from Telegram with /new.
Ask without interrupting. /btw did you update the docs too? gets an answer from a copy of the conversation while the run keeps going.
Keep it running. Run tasks on a desktop or VM so they don’t stop when the laptop sleeps: agents on other machines.

On a schedule, or when CI fails

Recipes are saved prompts with {params} (wisp run release 0.24.1). Triggers start them on their own while wisp is open:

Trigger	Fires when
Schedule	a cron expression matches, e.g. `0 9 * * 1-5`
CI failed	a GitHub Actions run fails, optionally on one branch
Issue labelled	an open issue carries a label

GitHub triggers poll with your gh login, so nothing has to reach your machine. Issue triggers start in Plan mode, since anyone who can open an issue writes that prompt.

Scripts and CI can use the CLI with --json, the JS SDK, the HTTP API or webhooks. See integrations.

Questions

Is Full auto safe?: Each task runs in its own git worktree, so your checkout is untouched, and every turn is checkpointed so you can roll back. For anything you are unsure about, use Plan (read-only) or Edits, where anything else the agent wants to run comes to you first.
Can I watch what it is doing?: Yes. A working card shows a live one-line status, the full transcript is one click (or V) away, and T opens the session in your own terminal.
Does it need to stay open?: wisp runs the agents, so it has to be running somewhere. Run it on a desktop or a VM through Machines and your laptop can sleep.

What plain headless runs miss

Start one

While you’re away

On a schedule, or when CI fails

Questions

Related guides