Deep Dive into Claude Code on the Web Specifications

Posted on November 1, 2025

Investigation results of the Claude Code on the Web environment as of November 1, 2025
Claude Code version was 2.0.25 at the time of writing
Sandbox behavior has already changed. Please test actual behavior in your own environment

Update: January 8, 2026 I updated this article’s data for the following article:

Using the gh Command in Claude Code on the Web:

Claude Code on the Webでghコマンドを使う

zenn.dev

Change summary:

Claude Code 2.0.25 -> 2.0.59
Updated basic environment specs
Skills added to allow permission
sonnet 4.5 -> opus 4.5
gh unlocked. System prompt changed

If you’re interested in the November specs, please check the change history in my zenn github repository.

I’m Oikon. I usually play around with AI tools, especially Claude Code.

Claude Code has Claude Code on the Web, an environment where you can run Claude Code in your browser. You can develop as long as you have a browser, even without a PC.

Claude Code on the web - Claude Code Docs

Run Claude Code tasks asynchronously on secure cloud infrastructure

docs.claude.com

Recently, I attended an event called aimeetup and was given the topic “Create something using Claude Code on the Web”, so I decided to investigate Claude Code on the Web specifications.

I hope this helps those who want to understand Sandbox environment limitations or master Claude Code on the Web.

What is Claude Code on the Web?

Claude Code on the Web is an environment for running Claude Code in your browser provided by Anthropic.

https://claude.com/code

claude.com

Unlike local Claude Code, the biggest appeal is that you can use it from anywhere as long as you have a browser. You can also launch it from the smartphone mobile app.

Basic usage is already written in the following article:

Claude Code on the webを触ってみた

zenn.dev

Sandbox Environment

Claude Code on the Web runs in a gVisor-based container environment. gVisor is an application kernel developed by Google, a Sandbox technology that enhances container security.

Basic Specs

Item	Specification
OS	Ubuntu 24.04.3 LTS
Container Runtime	gVisor (runsc)
CPU	16 cores
Memory	21GB
Disk	30GB
File System	9p protocol
Network	Via proxy (JWT authentication)

How Claude Code on the Web Works

overview

Claude Code on the Web clones a GitHub repository into a gVisor-based container environment and runs Claude Code. Here’s the flow:

Launch Claude Code on the Web
Select repository and start chat
Start Sandbox environment
Clone repository to Sandbox’s /home/user/
Launch Claude Code within git repository
Start session within launched Claude Code
Feed session content back to Claude Code on the Web
Continue session thereafter

In other words, Claude Code on the Web is technology that clones a git repository into a Linux environment and launches claude code. Understanding this premise reveals what’s possible in the Sandbox environment.

Claude Code on the Web Specifications

Since Claude Code on the Web clones a git repository, the only settings you can bring are project settings. That means you can’t bring ~/.claude/ global settings, but you can bring ./.claude/ project settings.

Claude Code on the Web Global Settings

In Claude Code on the Web, there are global settings for Claude Code in ~/.claude/settings.json within the Sandbox.

Contents are as follows:

settings.json

{
  "$schema": "https://json.schemastore.org/claude-code-settings.json",
  "hooks": {
    "Stop": [
      {
        "matcher": "",
        "hooks": [
          {
            "type": "command",
            "command": "~/.claude/stop-hook-git-check.sh"
          }
        ]
      }
    ]
  },
  "permissions": {
    "allow": ["Skill"]
  }
}

As you can see above, Stop Hooks and permissions (Skill permission) are configured. For those curious about the contents of stop-hook-git-check.sh, I’ll attach it here:

stop-hook-git-check.sh

#!/bin/bash

# Read the JSON input from stdin
input=$(cat)

# Check if stop hook is already active (recursion prevention)
stop_hook_active=$(echo "$input" | jq -r '.stop_hook_active')
if [[ "$stop_hook_active" = "true" ]]; then
  exit 0
fi

# Check if we're in a git repository - bail if not
if ! git rev-parse --git-dir >/dev/null 2>&1; then
  exit 0
fi

# Check for uncommitted changes (both staged and unstaged)
if ! git diff --quiet || ! git diff --cached --quiet; then
  echo "There are uncommitted changes in the repository. Please commit and push these changes to the remote branch." >&2
  exit 2
fi

# Check for untracked files that might be important
untracked_files=$(git ls-files --others --exclude-standard)
if [[ -n "$untracked_files" ]]; then
  echo "There are untracked files in the repository. Please commit and push these changes to the remote branch." >&2
  exit 2
fi

current_branch=$(git branch --show-current)
if [[ -n "$current_branch" ]]; then
  if git rev-parse "origin/$current_branch" >/dev/null 2>&1; then
    # Branch exists on remote - compare against it
    unpushed=$(git rev-list "origin/$current_branch..HEAD" --count 2>/dev/null) || unpushed=0
    if [[ "$unpushed" -gt 0 ]]; then
      echo "There are $unpushed unpushed commit(s) on branch '$current_branch'. Please push these changes to the remote repository." >&2
      exit 2
    fi
  else
    # Branch doesn't exist on remote - compare against default branch
    unpushed=$(git rev-list "origin/HEAD..HEAD" --count 2>/dev/null) || unpushed=0
    if [[ "$unpushed" -gt 0 ]]; then
      echo "Branch '$current_branch' has $unpushed unpushed commit(s) and no remote branch. Please push these changes to the remote repository." >&2
      exit 2
    fi
  fi
fi

exit 0

A brief explanation of stop-hook-git-check.sh operation:

Check to prevent Hooks looping
Verify Git repository
Check for commit changes
Check for untracked files
Check for unpushed commits

In other words, it’s prompting Claude to commit + push if there are differences in the repository. This is why Claude Code on the Web commits + pushes immediately after making changes - this Hook is configured.

Claude Code Startup Options

In Claude Code on the Web, Claude Code is launched within the Sandbox. The startup options configured for Claude Code are:

claude \
  --output-format=stream-json \
  --verbose \
  --replay-user-messages \
  --input-format=stream-json \
  --debug-to-stderr \
  --allowed-tools Task,Bash,Glob,Grep,Read,Edit,MultiEdit,Write,NotebookEdit,WebFetch,TodoWrite,WebSearch,BashOutput,KillBash,Tmux,mcp__codesign__sign_file \
  --disallowed-tools Bash(gh:*) \
  --append-system-prompt "You are Claude, ..." \
  --model claude-opus-4-5-20251101 \
  --add-dir /home/user/repo_name \
  --sdk-url wss://api.anthropic.com/v1/session_ingress/ws/session_id \
  --resume=https://api.anthropic.com/v1/session_ingress/session/session_id \
  --debug

Because the --verbose option is specified, startup options can be obtained by having Claude execute something like grep -A 5 "Executing Claude Code" /tmp/claude-code.log.

As you can see above, the model is fixed to Opus 4.5. Specifying “model” in settings.json won’t change it.

Startup options:

CLI reference - Claude Code Docs

Complete reference for Claude Code command-line interface, including commands and flags.

docs.claude.com

The gh command was previously in the prohibited tools, but has now been unlocked. For details, see the following article:

Claude Code on the Webでghコマンドを使う

zenn.dev

--append-system-prompt contents


You are Claude, an AI assistant designed to help with GitHub issues and pull requests. Think carefully as you analyze the context and respond appropriately. Here's the context for your current task:
Your task is to complete the request described in the task description.

Instructions:
1. For questions: Research the codebase and provide a detailed answer
2. For implementations: Make the requested changes, commit, and push

## Git Development Branch Requirements

You are working on the following feature branches:

 **repository-name**: Develop on branch `branch-name`

### Important Instructions:

1. **DEVELOP** all your changes on the designated branch above
2. **COMMIT** your work with clear, descriptive commit messages
3. **PUSH** to the specified branch when your changes are complete
4. **CREATE** the branch locally if it doesn't exist yet
5. **NEVER** push to a different branch without explicit permission

Remember: All development and final pushes should go to the branches specified above.

## Git Operations

Follow these practices for git:

**For git push:**
- Always use git push -u origin <branch-name>
- CRITICAL: the branch should start with 'claude/' and end with matching session id, otherwise push will fail with 403 http code.
- Only if push fails due to network errors retry up to 4 times with exponential backoff (2s, 4s, 8s, 16s)
- Example retry logic: try push, wait 2s if failed, try again, wait 4s if failed, try again, etc.

**For git fetch/pull:**
- Prefer fetching specific branches: git fetch origin <branch-name>
- If network failures occur, retry up to 4 times with exponential backoff (2s, 4s, 8s, 16s)
- For pulls use: git pull origin <branch-name>

The previous system prompt about gh prohibition has been removed.

The GitHub CLI (gh) is not available in this environment. For GitHub issues ask the user to provide the necessary information directly.

Features Available in Claude Code on the Web

Claude Code on the Web just launches Claude Code within a Sandbox, but some features from local Claude Code aren’t available. I’ll introduce what you can and can’t do.

CLAUDE.md

claudemd

CLAUDE.md contents are reflected. For example, if you write something like this:

## Context Window Report

**Important**: At the end of each response, briefly report context window usage.

Format:
/```
---
Context: Used XXK / Remaining XXK (XX%)
/```

Example:
/```
---
Context: Used 32K / Remaining 168K (84% remaining)
/```

Writing this confirmed that instructions are correctly reflected in Claude Code on the Web as shown in the image.

Slash Command

slashcommand

Slash commands (custom slash commands) cannot be used. This is because canUseTool is not available.

However, Claude is smart enough to indirectly execute slash commands by directly reading the Markdown file.

In the image, I’m writing in natural language like run /xyz. Currently, if you try to directly invoke /xyz as a command, Claude Code on the Web has a bug where no response is returned.

Hooks

Hooks can be used. As mentioned earlier, you can see that stop-hooks-check-git is configured in Claude Code’s global settings.

Subagents

subagent

Subagents can be used. The image above shows parallel execution of the default @agent-Explore subagent.

Unlike local Claude Code, Subagents are executed with the label Task, and specific Subagent names are not displayed. (This was also the case in Claude Code 1.0.x era)