Claude Code on the web vs Roost
Claude Code on the web runs one agent on machines Anthropic manages; Roost runs any CLI in a real terminal on macOS and Linux machines you own.
Pick Roost if
You want your code to stay on your own macOS or Linux hardware and any CLI, not just Claude Code, to run in a terminal you can open from a phone.
Pick Claude Code on the web if
You want zero infrastructure and are happy for Claude Code alone to run on someone else's VM.
Side by side
Scroll the table sideways on a phone.
| Feature | Roost | Claude Code on the web |
|---|---|---|
| Host platforms | macOS and Linux (including Linux arm64) | Anthropic-managed cloud VMs; enterprises may route sessions to self-hosted infrastructure; `claude rc` tethers one local machine |
| Client devices | any browser: desktop, iPhone, Android, iPad | any browser, plus the Claude iOS and Android apps |
| Multi-machine fleet | Yes | No |
| Zero-install client | Yes | Yes |
| Persistent sessions | Yes | Partial |
| Any CLI in a real PTY | Yes | No |
| Mobile UX | full app, touch selection, key row, swipe deck | the Claude mobile apps drive sessions that are already running |
| Voice input | Yes | No |
| Agent state + push | Web Push + working/needs-input/done | Session progress in the Claude web and mobile UI |
| Self-hosted, no account | Yes | No |
Where they differ
-
Who owns the machines. Anthropic runs the control plane and the VMs — or tethers one local machine with
claude rc. Roost is self-hosted only: terminal workers stay on your macOS and Linux machines and you operate the control plane yourself. There is no Roost-hosted tier and no Roost account. -
A task view versus the real terminal. The web product gives you a task-shaped view of what Claude Code did, and what runs in it is Claude Code. Roost gives you the real terminal, with full ANSI, scrollback, mouse, and touch, and what runs in it is any agent, shell, REPL, or TUI — including Claude Code, which Roost treats as an ordinary command.
-
“New terminal on demand” is the sharp edge. Away from your desk, the web product can only continue what already exists in a sandbox or on the tethered machine. Roost lets you browse folders on any enrolled worker from a phone and choose Open terminal here, on a machine you are nowhere near.
-
Session lifetime is bounded by different things. A managed sandbox lives as long as the session Anthropic is running for you, and the
claude rctether lasts as long as that one machine stays awake and connected. In Roost a keeper subprocess hosts every PTY and outlives worker restarts and updates, so closing the lid, losing WiFi, or updating the fleet withroost pushdoes not end the session — reattach later from a different device and the full scrollback is still there. -
Credentials and client identity. The sandbox is deliberately locked down: network egress goes through an allowlist proxy, and the GitHub token lives in a separate proxy outside the sandbox so the agent never holds the credential; your client identity is your Anthropic account session. Roost adds no credential broker — it uses whatever is already on the machine — and each device holds a per-device Ed25519 key minted in the browser, non-extractable, revocable by deleting a row.
-
Where the code lives. Claude Code runs in a cloud VM or on one tethered machine. Roost’s released product keeps the shell and code on workers you connect: it is self-hosted, accountless, and has no telemetry. Enterprises can route Claude web sessions to self-hosted infrastructure instead of Anthropic’s.
What you give up either way
- Choosing Roost costs you: managed execution sandboxes and most integration surfaces. You still provide and enroll the machines where terminals run; you also operate a coordinator, which binds loopback and sits behind a front door you choose — Caddy, a Cloudflare tunnel,
tailscale serve. Roost has one client, the browser, with no editor plugin, CI action, or chat bot, while Anthropic maintains terminal, web, iOS, Android, VS Code, JetBrains, GitHub Actions, and Slack paths. - Choosing Claude Code on the web costs you: a shell, panes, and scrollback; more than one machine; opening a fresh terminal while away; any CLI other than Claude Code; and code that never leaves hardware you control.
Use Claude Code on the web instead if…
You want managed cloud execution and are happy for Claude Code alone to run on someone else’s VM. If your work fits in a sandbox with an allowlisted network, Anthropic operates the isolation, credential proxying, and machine. Roost is not a replacement for that: its released product is self-hosted and supplies neither execution sandboxes nor agent machines, and it wins only when the code must stay on your hardware, the tool is not Claude Code, or the work needs a real terminal.
What Claude Code on the web is
Claude Code on the web runs Claude Code sessions in isolated Anthropic-managed VMs, driven from the browser or the Claude mobile apps. It is the same Claude Code you would run in a terminal, with the terminal removed: you get a task-shaped view of what the agent did, not a shell. There is nothing to install and nothing to operate, which is the entire point.
Sources
- Claude Code (claude.com/product/claude-code)
- Claude Code on the web (claude.ai/code)
See the whole field on the alternatives hub, or read how device keys work in security.