Claude Code on the web runs Claude Code in Anthropic’s cloud infrastructure against a clone of your repository, so it works without your machine. A local agent with a room runs Claude Code on your own computer and lets people you invite send it work while you approve what runs. The first is always on; the second is always yours.
Both are legitimate. They differ on where the code and credentials live, who can talk to the agent, and what happens when nobody is watching.
How Claude Code on the web works
You connect a repository, Anthropic provisions a sandbox, clones the code and runs Claude Code there. Git operations use tokens issued to the provider through GitHub’s OAuth integration; your local credentials, SSH keys and environment variables are not involved. Results come back as branches or pull requests. The session is yours, from any browser.
This is the right shape for unattended background work: a queue of small tasks across many repositories, running while your laptop is closed.
How a local agent with a room works
Claude Code runs on your machine as it always has, with the repository you already have checked out and the credentials already there. The muti daemon polls for work over an outbound connection and streams each tool call into a room. You, or anyone you invite, send requests from a browser or a phone. Requests queue while the machine sleeps. Dangerous actions pause for you.
This is the right shape for sharing: several people, one agent, your terms.
Provider cloud on the left, your machine on the right, and what each is good for.
Side by side
Where each one is stronger
Web wins when the machine cannot be relied on: overnight batch fixes across a dozen repositories, contributors without a development environment, or a compliance posture that prefers developer laptops never to run the agent at all.
Local wins when the agent needs what is on your machine: a database only reachable from your network, a design tool logged in on your browser, a monorepo too large to clone per task, or MCP servers you have configured. It also wins whenever a second person is involved, because a room gives that person a way to ask and gives you a way to say no.
The credential question
The web version’s security argument is real: your local secrets are never in the agent’s environment. The local version’s argument is the mirror image: your secrets are never in anyone else’s environment. Which matters more depends on whom you trust less, your own laptop or a provider’s multi-tenant sandbox. Rooms add one thing the web version cannot: when someone else asks, a credential prompt routes to the owner, and the requester never sees it. See what the daemon uploads.
Decision table
Frequently asked questions
Is Claude Code on the web the same as Remote Control? No. Remote Control drives a session on your machine from another device; the web version runs in the cloud. See Remote Control vs an agent room.
Can a room use Claude Code on the web as the agent? No. A room needs an agent on a machine the owner controls; that is what makes the approval model hold.
Does the local version need an open port? No. The daemon polls outbound. Nothing connects in.
Keep the agent where the work is. If the work is on your machine, give the agent a room.
Published