How to Use Claude Code on the Web

Claude Code accounted for approximately 4% of all public GitHub commits by February 2026, with projections reaching 20% by year-end. Nearly every developer tutorial describes it as a terminal tool you install locally, run against your codebase, and close when you're done.

That framing misses the other half. Since October 2025, Claude Code has run in the browser at claude.ai/code. The web interface executes tasks on Anthropic-managed virtual machines, which means no local environment setup, no open terminal window, and sessions that keep running after you close your laptop. It's available as a research preview for Pro, Max, Team, and Enterprise subscribers.

The web version introduces capabilities that don't exist in the terminal. Auto-fix monitors your pull requests and pushes fixes when CI breaks or a reviewer leaves a comment. Session teleportation lets you start work in the browser and pull it into your terminal later, or send local tasks to the cloud and check results from your phone. You can run multiple sessions in parallel, each on a different repository and branch, without juggling local worktrees.

This matters for teams where both humans and AI contribute to the same projects. A developer can submit a well-defined task from their phone during a commute, monitor progress from a browser at their desk, then teleport the session into their terminal for final adjustments. The work stays continuous across every surface.

The rest of this guide walks through the full workflow: connecting GitHub, configuring cloud environments, submitting tasks, and using the features that make the web interface worth adding to your development routine.

Getting started takes about five minutes. You need a GitHub account and a Claude Pro ($20/month), Max ($100 or $200/month), Team, or Enterprise subscription.

Step 1: Sign in and connect GitHub

Go to claude.ai/code and sign in with your Anthropic account. The interface prompts you to install the Claude GitHub App. Follow the authorization flow to grant access to your repositories. Cloud sessions can access any repository your GitHub account can see, regardless of which repos you install the App on specifically.

If you already use the GitHub CLI, skip the browser flow. From the Claude Code terminal, run:

/web-setup

This syncs your local gh token to your Claude account and creates a default cloud environment.

Step 2: Create your cloud environment

After connecting GitHub, you configure a cloud environment that controls what runs when sessions start and what network access Claude gets. The form has four fields:

• Name: A label to distinguish environments, useful when you have different setups for different projects
• Network access: Controls outbound connections. The default "Trusted" level allows common package registries (npm, PyPI, Docker Hub) while blocking general internet access
• Environment variables: Optional key-value pairs in .env format, available in every session
• Setup script: An optional Bash script that runs before Claude Code launches, then gets cached so it doesn't re-run on subsequent sessions

For a first project, leave the defaults and click Create environment. You can edit this environment or add more later.

Step 3: Submit your first task

Select a repository and branch from the selector below the input box. Choose a permission mode: "Accept edits" lets Claude work autonomously, while "Plan mode" has Claude propose an approach and wait for approval before editing files.

Type a specific task description. Name the file or function:

• "Add a README with setup instructions for the Python SDK"
• "Fix the failing auth test in tests/test_auth.py"
• "Refactor the logger in src/utils/logger.ts to use structured output"

Claude clones the repo, runs your setup script if configured, and starts working. Each task gets its own isolated session and branch, so you can start another task immediately without waiting for the first to finish.

Each session runs on an isolated Anthropic-managed VM with 4 vCPUs, 16 GB of RAM, and 30 GB of disk. The VM comes pre-loaded with development tools across major languages:

• Python: Python 3.x with pip, poetry, uv, black, mypy, pytest, ruff
• Node.js: Versions 20, 21, and 22 via nvm, plus npm, yarn, pnpm, bun, eslint, prettier
• Ruby: 3.1, 3.2, and 3.3 with bundler and rbenv
• Java: OpenJDK 21 with Maven and Gradle
• Go and Rust: Latest stable toolchains with module and cargo support
• Docker: Full engine with compose, plus pre-allowed access to Docker Hub, GHCR, and ECR
• Databases: PostgreSQL 16 and Redis 7.0, installed but not running by default

Your repo's CLAUDE.md, .claude/settings.json, and .mcp.json files carry over automatically because they're part of the clone. User-level config from ~/.claude/ does not transfer. To add tools that aren't pre-installed, write a setup script. The result is cached for roughly seven days, so subsequent sessions start fast without re-running the script.

Network access levels

You control outbound connections per environment:

• None: No internet access at all
• Trusted: Allowlisted domains only, covering npm, PyPI, RubyGems, crates.io, GitHub, Docker Hub, and cloud SDKs from AWS, GCP, and Azure
• Full: Unrestricted internet access
• Custom: Your own domain allowlist, optionally merged with the Trusted defaults

GitHub operations always go through a dedicated proxy regardless of this setting. The proxy handles authentication using scoped credentials so your tokens never enter the container.

Running tests and starting services

Test runners like pytest, jest, and cargo work out of the box. Ask Claude to run them in your prompt ("run pytest after each change") or let Claude decide when to verify its work. PostgreSQL and Redis need to be started per-session since the cache preserves files, not running processes. Ask Claude to run service postgresql start or include the command in a SessionStart hook.

Docker compose works for spinning up multi-container stacks. If your images are large, add docker compose pull to your setup script so they're cached and ready at session start.

Store and share what your cloud sessions produce

A workspace with built-in RAG and MCP access. Give your AI coding workflows a persistent home for artifacts that don't belong in a GitHub repo. Starts with a 14-day free trial.

Start 14-Day Trial

The ability to move work between surfaces is what separates Claude Code on the web from a basic cloud IDE. Two commands handle it: --remote sends work to the cloud, and --teleport pulls cloud sessions back to your machine.

Send work to the cloud with --remote

Start a cloud session from your terminal:

claude --remote "Fix the authentication bug in src/auth/login.ts"

This creates a new session at claude.ai/code, clones your current repo at its current branch, and starts working. Push local commits first, since the VM clones from GitHub rather than your machine. The task runs in the cloud while you continue working locally.

Run multiple --remote commands to execute tasks in parallel:

claude --remote "Fix the flaky test in auth.spec.ts"
claude --remote "Update the API documentation"
claude --remote "Add input validation to the signup endpoint"

Each creates its own session with its own branch. Monitor them with /tasks in the CLI, or check progress from the web or Claude mobile app.

For repos not on GitHub, Claude Code bundles your local repository (up to 100 MB) and uploads it directly to the cloud session. This happens automatically when no GitHub remote is detected.

Pull cloud sessions locally with --teleport

When you want to continue a cloud session with your full local tooling:

claude --teleport

This opens an interactive picker showing your cloud sessions. Select one, and Claude verifies you're in the right repo, fetches the branch, checks it out, and loads the full conversation history into your terminal. Your working directory must be clean (no uncommitted changes) and you need to be in a checkout of the same repository.

You can also teleport from inside a running CLI session with /teleport (or /tp), or from the /tasks list by pressing t.

Pre-fill sessions with URL parameters

Build integrations by constructing URLs with query parameters:

https://claude.ai/code?prompt=Fix%20the%20login%20bug&repositories=acme/webapp

This opens claude.ai/code with the prompt and repository pre-selected. Use it to link from issue trackers, Slack bots, or internal dashboards so your team can launch targeted coding sessions in one click.

Auto-fix is the feature that makes the web interface pull its weight for team workflows. Once enabled on a pull request, Claude watches for CI failures and reviewer comments, then investigates and pushes fixes without you re-opening the session.

Enabling auto-fix

There are several ways to turn it on depending on where the PR came from:

• From the web: Open the CI status bar on a session-created PR and toggle Auto-fix
• From terminal: Run /autofix-pr while on the PR's branch
• From mobile: Tell Claude to "watch this PR and fix any CI failures or review comments"
• For any existing PR: Paste the PR URL into a session and ask Claude to auto-fix it

Auto-fix requires the Claude GitHub App installed on the repository. If you connected via /web-setup and want auto-fix later, install the App on those repos through GitHub's Settings page.

How Claude responds to PR activity

When a CI check fails or a reviewer leaves a comment, Claude reads the output and decides how to respond. For clear fixes, like a missing import, a typo, or a test assertion off by one, Claude makes the change, pushes it, and explains what it did in the session. For ambiguous requests or decisions that could affect the project's architecture, Claude asks you first. Duplicate or no-action events get noted and skipped.

Claude may reply to review comment threads on GitHub as part of resolving them. These replies post under your GitHub username but are labeled as coming from Claude Code, so reviewers know an agent wrote the response rather than you.

One limitation worth knowing: GitHub doesn't send webhooks when the base branch advances and creates a merge conflict. You need to open the session and ask Claude to rebase manually.

Auto-fix is a per-PR toggle. Turn it off from the CI status bar or tell Claude to stop watching the PR.

Claude Code on the web handles code changes through GitHub branches and pull requests. But many AI-assisted workflows produce output that doesn't belong in a repository: generated reports, processed data files, client deliverables, design specifications, or documentation packages for handoff.

Local file storage works for single-developer projects but breaks down when agents need to hand off work to humans or other tools. S3 and Google Cloud Storage handle raw persistence if you're willing to manage buckets, IAM policies, and signed URLs. Google Drive and Dropbox offer simpler sharing but weren't built for workflows where files are created programmatically and need structured organization.

Fast.io addresses this gap as a workspace where both agents and humans operate on the same files. Plans start with a 14-day free trial, with workspaces, monthly AI credits, and storage scaled to each tier. Agents connect through the Fast.io MCP server using Streamable HTTP at /mcp, which exposes 19 tools for file operations, AI queries, workspace management, and share creation.

A practical workflow looks like this: Claude Code writes and commits source code to GitHub through its normal PR flow. Generated artifacts like documentation, test reports, or processed datasets get uploaded to a Fast.io workspace via MCP. Intelligence Mode auto-indexes those uploads for semantic search and RAG queries. When the project is ready for handoff, workspace ownership transfers from the agent to a human, who can create branded portals for clients or share links with teammates.

This separation keeps your GitHub repos focused on code while giving non-code output a searchable home that both AI tools and humans can access. For teams running multiple Claude Code sessions in parallel across different projects, having a central workspace for generated artifacts prevents files from scattering across local machines, Slack threads, and email attachments.

Read the agent onboarding guide or try the MCP server docs to connect your first agent workflow.