AL@alicedriving
BO@bobtyping
CA@carolin stats.py
Live presence
See who is in the session, who is typing and which file each teammate has open.
Early accessFree while we build it with the first teams
One coding agent. Your whole team in the room.
huddle runs an AI coding agent on your repository in a shared session. Everyone sees each step as it happens, one person drives, anyone can suggest, and risky commands wait for a human to say yes.
How it works
A session is a shared room around one agent. Not a chat you paste into Slack afterwards.
-
Start from a repo and a task
Pick a GitHub repository and branch and say what done looks like. The agent gets its own sandbox and a fresh checkout.
-
Invite the people who know the code
Share the link. Teammates join live and see every command, file read and diff, and can comment on any step.
-
Steer it together
The driver talks to the agent. Everyone else suggests, and the driver sends, edits or drops each suggestion.
-
Approve, then hand off
Pushes, deletes and network calls wait for a person. Leaving? Pass the wheel with a written summary.
Built for teams
Everything the agent does, where everyone can see it.
ALBOCA
One driver, many hands
One person instructs the agent so it never gets two plans at once. Anyone can ask for the wheel.
git push origin HEADwaiting for approvalapproved by @carol
Approvals
Risky commands stop and ask. The session records who said yes.
def mean(values):
- return sum(v) / (len(v) - 1)
+ return sum(v) / len(v)
Diffs as they happen
Review changes while the agent writes them, or fix a line yourself and commit.
network: offsandbox
Isolated sandbox
A container per session, no network unless a person approves it.
The product
The repository, not a transcript of it.
Every step the agent takes, live, with the controls the driver needs on the right.
Guardrails
The agent can do a lot. Some of it waits for you.
The sandbox has no network by default. These actions stop and ask, and the session shows who approved what and when.
| Action | Example | What happens |
|---|---|---|
| git_push | git push origin HEAD | waits for approval |
| delete_files | rm -rf build/ | waits for approval |
| network | pip install requests | waits for approvalthen runs with network for that one command |
| write_outside_workspace | echo x > /etc/hosts | waits for approval |
| everything else | pytest -q | runsin the sandbox, visible to everyone |
Questions
Asked so far
Which model does the agent use?
Claude, through Anthropic's API. Each session shows the model it runs on and how many tokens it has used.
Where does my code run?
In an isolated container per session, created from your repository and removed when the session ends. It has no network unless someone approves a command that needs it.
Can two people type to the agent at once?
No, on purpose. One driver at a time keeps the agent from getting contradicting instructions. Everyone else's suggestions queue up for the driver.
Does it push to my GitHub?
Only after a person in the session approves that exact push command. The approval, and who gave it, stay in the session's history.
What does it cost?
Nothing during early access. We will announce pricing before it changes, and existing teams keep their sessions.
How do I get access?
Anyone can sign in and join a session they're invited to. To start your own sessions, write to us and tell us a little about your team.
Bring the agent into the room
Sign in with GitHub, open a session, send the link to one teammate.
Start with GitHub