# Run a live session (/tutorials/live-sessions)

# Run a live session [#run-a-live-session]

A checklist run or an interview round goes off and comes back with a report. A
live session is the opposite: you sit beside the actor while it works. The
browser is on the left, the chat on the right, and underneath it the steps the
actor took and the captures it kept. Ask it to try something, watch what it
does, change the task halfway through, or take the browser and drive it
yourself.

<Callout variant="note" title="Early Access">
  Live sessions are in early access and switched on per workspace. When they are on for
  yours, **Live sessions** appears under **Discover** in each project's
  navigation.
</Callout>

## Start a session [#start-a-session]

<Steps>
<Step>

### Open Live sessions [#open-live-sessions]

Open a project and choose **Live sessions** under **Discover**. The page lists
the project's sessions, newest first, so you can go back to any earlier one and
read what happened. A session that has ended opens on its summary. The latest
few also appear on the project's overview.

</Step>
<Step>

### Click New session [#click-new-session]

The dialog asks only what cannot change once the session is running:

- **Actor**: who uses the product.
- **Start on**: the address the browser opens first.
- **Session**: **Interview** or **Task** (see below).
- **Browser**: **This Mac** or **Cloud browser**.
- **Model**: **Stunt Double**, or a model on this Mac. Asked only when the
  browser is This Mac and the Mac already has
  [local models](/integrations/desktop-app#local-models).

Replies, tools and sign-ins are all changed from inside the session.

</Step>
<Step>

### Click Start session [#click-start-session]

The browser opens on the start address and the actor is ready for its first
message.

</Step>
</Steps>

## Interview or task [#interview-or-task]

- **Interview** (the default): the actor stays in character and thinks aloud
  while you moderate. Ask it what it expects a page to do, what it would click
  next, or what would stop it signing up. This is a moderated usability
  session without recruiting anyone.
- **Task**: the actor does what you ask and reports back. Use it to reproduce
  a bug, try a flow end to end, or check something quickly before a release.

## Choose a browser [#choose-a-browser]

- **Cloud browser** works from anywhere, on public sites.
- **This Mac** is offered when you are in the
  [desktop app](/integrations/desktop-app) and your plan includes devices. The
  actor drives a Chrome on your Mac, so it can open local dev servers, staging
  behind your VPN and anything else this Mac can reach. When a session can run
  on this Mac, it is the default.

A project that is only reachable inside your network needs **This Mac**: the
cloud browser refuses it before the session starts. Private and internal
addresses also need the Startup plan or above.

On This Mac, the browser follows the app's
[Where the browser runs](/integrations/desktop-app#how-the-browser-runs)
setting: inside the app by default, where you watch it in the session, or in
its own Chrome window.

You do not need to set your Mac up as a device first. If it is not one yet,
starting the session pairs it for you in the background, and that pairing is
removed again when its last session ends. A Mac that already belongs to
someone else in the workspace, or to another workspace, is never moved: the
dialog says so and offers the cloud browser instead.

## Steer the actor [#steer-the-actor]

Type in the chat to give the actor a task or ask it a question. While it is
working, a new message is queued for when it finishes. Choose **Send now** to
steer the current task instead of waiting.

Each step it takes appears in the journey under the browser, grouped by task,
so you can see exactly what it clicked, typed and read.

### Replies [#replies]

The actor keeps replies short by default: the result first, no recap of the
task and no commentary on each click, since you can see the browser yourself.
Open **Replies** to change that from its next reply:

- **Detail**: from a sentence to a thorough account with evidence.
- **Narration**: from results only to thinking aloud at every step. An
  interview starts with more narration than a task.

## Take control [#take-control]

Click **Take control** in the header, or click or type in the browser
itself. If the actor is in the
middle of a step, it stops first. Now the browser is yours: sign in, fill a
form, dismiss a banner or get the page into the state you want.

Click **Return control** when you are done. The actor is told you drove the
browser in the meantime, so its next step starts from the page as you left it.

## Keep what it finds [#keep-what-it-finds]

- **Findings** are recorded by the actor as it works, with a severity and the
  screenshot that shows the problem.
- **Capture** saves a screenshot of the browser as it is now. It is off while
  the actor is driving.
- **Notes** are your own, added from the captures tray.

Captures stay with the session after it ends.

## Tools and sign-ins [#tools-and-sign-ins]

Open **Tools** during a session to see what the actor can use beyond the
browser.

- **Project and actor** MCP servers you have already connected in Stunt Double.
- **On this Mac**: MCP servers that run on your own Mac, in sessions using the
  desktop app. Turn them on for the session, then import servers from Claude
  Desktop or Cursor, or add them in the app (see
  [MCP servers on this Mac](/integrations/desktop-app#mcp-servers-on-this-mac)
  for how the app approves them). The chat asks you before each tool call.
  Switch on **Run read-only tools without asking** to skip that question for
  tools their server marks read-only.

What a tool returns is kept in the session's transcript, which everyone in the
workspace can read, so only turn on servers whose output you are happy to
share.

**Browser sign-ins** (sessions on This Mac) controls where the browser keeps
its cookies and sign-ins:

- **Per actor**: each actor keeps its own sign-ins, the way two real people
  would. This is the default.
- **Shared in this workspace**: sign in once and every actor in the workspace
  starts signed in.

Sign-ins stay on your Mac, in a browser profile kept apart from your own
Chrome, so an actor never sees your personal sessions. The same section lists
the profiles stored on this Mac with their size, and can clear one.

A Mac MCP server that starts with `npx` or `uvx` needs Node or uv, which
[Set Up This Mac](/integrations/desktop-app#set-up-this-mac) installs.

## End the session [#end-the-session]

Click **End session** and confirm. The browser closes and the transcript,
journey and captures stay where they are.

Once a session ends, a summary takes the browser's place: how long it ran,
the pages visited, the findings by severity and your notes, with a short
account of what happened and suggested next steps. A session that thought with
a model on your Mac gets the facts without the written account, since its
transcript is not sent to Stunt Double's model. Opening an ended session from
the list, or from the actor's activity, takes you straight to its summary.

A session left idle closes its browser on its own after a while. When that is
close, **Keep session open** appears in the header with the time it closes.

## Who can see a session [#who-can-see-a-session]

Only the person who started a session can chat with the actor or drive the
browser. Everyone else in the workspace can open it and follow along: the
journey, the captures and the transcript update live, and the browser pane
shows the latest screenshot.

## Next steps [#next-steps]

- **[Set up the desktop app](/integrations/desktop-app)** to run sessions on your Mac
- **[Run an interview](/tutorials/running-an-interview)** across a whole panel of actors
- **[Write effective checklists](/tutorials/effective-checklists)** to keep checking what a session found