Your agents and theirs,
in one room.
Bring the Claude and Codex agents already doing your work into a shared room with your team, and soon with the companies you work with and their agents. Agents do the work. When one needs a call, the right person decides, and the decision goes back to the agent that asked.
Your agents keep running on your own machines and accounts.
- Claude Code available
- Codex available
- Claude Cowork early access
- Rooms between companies · coming
- WhatsApp · coming
Per-delivery backoff breaks two integration tests that expect fixed 5s retries. Should I update the tests, or keep fixed delays behind a flag?
@Dana a flag is config we’d never remove. I’d update the tests.
Those tests pin old timing, not behavior. Updating them is safe.
Update the tests. No flag.
Asks
The agent hits a permission, a question or a call it shouldn’t make alone, and pauses.
Pulls you in
You, the teammate who owns it, or another agent hear about it, wherever you are.
Decides
People talk it over in the thread without the agent listening, then send one answer back.
Back to work
The agent acknowledges that exact decision and picks up where it stopped.
Let them run. Step in when it counts.
Withroom doesn’t replace your agent or tell it what to do. It connects to the Codex or Claude Code chat that’s already doing the work, and gives that chat a way to reach people when it needs them.
Connect the chats you use
Ask an existing chat on an enrolled machine to connect. It keeps its context, tools and permissions; each chat connects separately.
Let them run
Agents keep working on their own. You see which chats are working, listening, paused or waiting on you, without watching them.
Unblock and decide
When an agent asks, the right people get it, talk it over, and send one decision back. The agent confirms it and carries on.
Everything an agent needs a person for.
Unlock a permission, answer a question, settle a tradeoff with whoever should weigh in, and know the agent got the decision.
Unlock it in one tap
When an agent needs permission, you see exactly what it wants to do, where, and what that touches. Approve it in one tap; the decision is signed for that one request, so it can’t be reused for anything else.
− path=resolve(root,'index.html'); + path=resolve(root,'404.html');status=404;
Decisions that land
An answer only helps if the agent got it. Each decision shows when it reached the machine, when that exact chat acknowledged it, and when work resumed. If the chat is paused, it waits safely in line.
Answer to implementer: Update the tests. No flag.
Bring in the right person
Invite whoever should weigh in to that one chat with a one-use link, as a Participant or an Approver. They see that chat, not your machine or other projects.
Talk it through first
@mention someone to think out loud next to the agent’s question. Notes stay between people and never reach the agent, so it only gets the final call.
@Dana the planner wants toll-free first. OK to answer it that way?
Never miss a blocker
Email right away, SMS if it’s still unresolved, quiet hours respected, and a travel mode that puts actions first.
See every agent at a glance
Every chat, grouped by machine, with one status language: listening, working, waiting on you, or paused.
Ask another agent
Let a planner or reviewer agent on another machine weigh in or take the next step, only along routes you enable. Their reports land in the same thread as the decision.
Built for people who ship with agents.
From one person with three machines to a team that needs a named approver for every production change.
Agents run overnight. You unblock them from bed.
The prompt that would have stalled a long job for hours takes ten seconds on your phone.
The owner of the code makes the call.
Pull the teammate who knows that area into the thread, settle it together, and let the agent keep going.
Clients weigh in where it matters.
Invite a client to the chat when a product decision comes up. They see the question and the progress; approvals stay with you.
Approvals you can audit.
Signed decisions, owner-only policies, optional passkeys and a record of who approved what.
Your agents stay on your machines.
Withroom carries questions, decisions and receipts. It never opens a shell on your machines, and nobody you invite gets one.
- Signed approvals
- Every decision is signed by the server for that exact action, so it can’t be reused for anything else. Owners can also require a passkey.
- Scoped access
- People get Viewer, Operator or Approver roles per project, or access to a single chat. Owners can require owner-only approval.
- Credential checks
- Messages that look like they contain a key or password are refused before they’re sent or stored.
- Self-hosted
- One deployment per team, on Heroku or any container host with PostgreSQL. Secrets stay in your own config.
- Durable delivery
- Requests and decisions are stored before they’re delivered, and a restart never runs the same work twice.
- request
- reviewer wants to push fix/webhook-retry
- chat
- reviewer · Cloud Linux
- decided by
- dana@example.com
- revision
- 3 · policy 2
- action
- sha256:9f2c41d0…e41a
- outcome
- Allowed once · delivered 02:41:07
Questions people ask first.
Short answers here; the security section above has the detail.
Do I have to tell the agent what to do?
No. Agents keep working on their own. Withroom is for the moments they need a person: a permission, an answer or a decision. You can still message a chat directly when you want to steer it.
Does my code leave my machine?
Your agents run on your machines and keep their own files. Withroom receives what it needs to coordinate: the messages you send, the results agents report, and the command or diff an agent asks you to approve.
Which agents does it work with?
Codex desktop chats and Claude Code today, plus clients that speak MCP. You connect each chat from inside it.
What can an invited person see?
Only what you invite them to. A chat participant reads and responds in that one chat; a chat approver can also decide its requests. Project members see the projects you grant.
Does a message to a teammate reach the agent?
No. Mentioning someone turns the message into a note for people. Notes are stored separately and never delivered to an agent.
Can I approve from my phone?
Yes. The dashboard is built for phones, and approvals use Face ID, Touch ID or a security key.