Claude Code on the Web: What It Is, How to Use It

Claude Code on the web is Anthropic's hosted, asynchronous version of Claude Code. You open claude.ai/code, connect a GitHub repo, describe a task, and Claude Code runs it on an Anthropic-managed cloud machine and opens a pull request. You can start the same kind of session from your terminal with claude --cloud. It is a research preview, it needs a paid Claude plan, and on the phone it is a way to watch a task run — not a terminal you type into. If what you actually want is an interactive Claude Code shell on your phone, that is a different path, and we cover it below.

We wrote this after using both sides ourselves — Claude Code on the web for fire-and-forget PRs, and a cloud container for the live-terminal work we do from a phone. Every fact below is checked against Anthropic's own docs, the Claude Code on the web documentation and the launch blog, both re-read on 2026-07-07. Where a number could drift, it carries that date.

What is Claude Code on the web?

It is the same Claude Code agent, running on Anthropic's infrastructure instead of your machine. Anthropic launched it on 2025-10-20 and it is still labelled a research preview as of 2026-07-07. The model is delegation: you point it at a GitHub repository, write a task, and it works autonomously and opens a pull request with a summary of what it changed. You are not driving a shell keystroke by keystroke; you hand off a job and check back on it.

• Where it runs: Anthropic-managed cloud infrastructure at claude.ai/code, or via claude --cloud "…" from the CLI.
• The environment: a fresh, isolated Ubuntu 24.04 VM per session — roughly 4 vCPUs, 16 GB RAM, and 30 GB of disk, with Docker, PostgreSQL 16, and common language toolchains pre-installed.
• Parallelism: every task is its own session, so you can run several at once across repos.
• GitHub-centric: it clones from GitHub and pushes results back, and a proxy restricts pushes to the current working branch for safety.

The design goal is clear once you use it: this is a way to spawn a coding agent that lives in the cloud, does a bounded piece of work, and produces a reviewable diff. It is genuinely good at that.

How do you use Claude Code on the web?

Four steps: connect GitHub, submit a task, monitor it, then review and merge or teleport it to your terminal.

Step 1: Connect your GitHub account

Open claude.ai/code on a paid Claude plan and authorize GitHub — either through the Claude GitHub App (which also enables auto-fix on PRs) or a scoped token. Cloud sessions need repo access to clone code and push branches. Your real token never enters the container; git operations run through a dedicated proxy.

Step 2: Submit a task

Pick a repository and describe a well-scoped change — "add input validation to the signup handler and a test for the empty-email case," say. Claude Code clones the repo into a fresh VM and starts working. From the terminal, the same thing looks like this:

Step 3: Monitor the run

Watch progress in the web session or from the Claude mobile app. You can steer Claude, answer its questions, and leave inline comments on the diff it is building. This is the part people misread: monitoring is not a terminal. Anthropic's docs are direct about it — interactive commands that open a picker, like /model or /config, are not available in cloud sessions, because there is no interactive shell behind the web view. You supervise; you do not type ls.

Step 4: Review, open a PR, or teleport

When the session finishes, review the diff (it shows a +42 -18-style indicator), open a pull request from the web, or pull the whole thing onto your own machine to keep working locally:

Teleport is one-way from the CLI (web → local), and it needs the Claude Code CLI signed in to the same claude.ai account. If you leave a session idle, the VM is reclaimed and you will see it marked expired — more on that next.

What Claude Code on the web does not do

None of this is a knock — it is a different shape than a phone terminal, and naming the edges keeps you from picking the wrong tool. As of 2026-07-07:

• It is not free. Research preview for Pro, Max, and Team, plus Enterprise premium seats. No free tier, and no API-key or BYOK path — it routes Anthropic's own models on your subscription.
• It runs Claude Code only. No Codex CLI, Gemini CLI, or OpenCode. One agent, Anthropic's models.
• The VM is ephemeral. The conversation persists, but the box does not. Anthropic's docs: cloud sessions stop after inactivity and the environment is reclaimed, surfacing as its environment has expired. Creating a fresh session instead.
• It is GitHub-centric. Clone and PR require GitHub; GitLab or Bitbucket work only via a local bundle, with no pushback.
• Mobile is monitoring, not a shell. Great for approving a diff from the couch; not a place to run npm test yourself.

If you want an interactive Claude Code terminal on your phone

Here is the opinion we will plant a flag on, and one Anthropic's design would push back on: for mobile coding, async delegation is the wrong default. Most of the time on a phone — on the train, on the couch, in a waiting room — you do not want to fire a task and wait for a PR notification. You want to watch the agent think, catch it going the wrong way on line three, and say "no, use the existing helper" while it is still working. That is a live terminal, not a task queue. Claude Code on the web is excellent for the delegate-and-review job; it is not built for the type-and-watch job.

The type-and-watch job is what we build Cosyra for. It is a native iOS and Android app connected to a persistent Ubuntu 24.04 container on Azure AKS, and we ship Claude Code, Codex CLI, OpenCode, and Gemini CLI pre-installed. You open the app, you are in a real shell, you run claude and it is an interactive session — the same one you would get on a laptop, reachable from your phone.

They are not really substitutes. We use Claude Code on the web to delegate a bounded fix and get a PR back; we use Cosyra when we want to sit in the shell and drive. The honest trade-off on our side: Cosyra is cloud-only too, so neither of us runs offline, and we do not offer a self-hosted option. If you need offline or on-prem, both of us are the wrong answer.

Which one should you reach for?

• Pick Claude Code on the web if you already pay for Pro, Max, or Team, your code is on GitHub, and your job is "delegate a scoped change and review a PR later." The async parallelism and auto-fix are hard to beat for that.
• Pick a cloud container (Cosyra) if you want to type into a live terminal from a phone — including Android — run more than one agent, bring your own model keys, and keep a workspace that survives between sessions. Our Cosyra vs Claude Code on the web comparison puts the two side by side on pricing, terminals, and persistence.
• Pick Remote Control if you want to drive a Claude Code session running on a machine you already own — see Cosyra vs Claude Remote Control for that split.

For the broader map of where each agent runs on a phone, the AI coding agents on mobile pillar lays out all four, and remote control or cloud walks through the relay-versus-container decision in depth. If you specifically want the interactive Claude Code setup, start with Claude Code on phone.

Frequently asked questions