Skip to content

Claude Code Web - Technical Architecture Overview

Overview

Claude Code Web is Anthropic's browser-based agentic coding assistant that allows developers to delegate coding tasks directly from their browser. Launched in November 2025 as a research preview, it represents a fundamental shift from traditional desktop IDE tools by executing all operations within secure, isolated cloud sandboxes on Anthropic's infrastructure.

Unlike the CLI version that runs locally, Claude Code Web provides:

  • Browser-based access - No terminal or local installation required
  • Parallel task execution - Run multiple coding jobs simultaneously
  • Secure sandbox isolation - gVisor-based kernel-level isolation
  • GitHub integration - Direct repository access via OAuth

High-Level Architecture

Core Architecture Components

1. Sandbox Execution Environment

Each Claude Code Web session creates an isolated container with:

2. Security Architecture - gVisor Isolation

Key Security Features:

FeatureImplementationPurpose
Kernel IsolationgVisor user-space kernelPrevent direct hardware access
Syscall FilteringSeccomp-BPF (70-80 of 300+ allowed)Block dangerous operations
Network ControlEgress proxy with JWT validationRestrict outbound connections
Filesystem IsolationRead/write only to working directoryPrevent file system escape
Credential ProtectionScoped credentials via proxyNever expose real tokens

3. GitHub Integration Flow

Git Proxy Security Model:

  • Real OAuth tokens never enter the sandbox
  • Scoped credentials are validated per-operation
  • Branch restrictions enforced at proxy level
  • All git interactions logged and auditable

4. Network Egress Architecture

Task Execution Flow

Parallel Task Execution

Parallel Execution Benefits:

  • Run multiple tasks across different repositories
  • No additional cost - pay for compute time, not concurrent sessions
  • Each sandbox is fully isolated
  • Rate limits consumed proportionally

Real-Time Communication

Communication Features:

  • Bidirectional WebSocket for real-time updates
  • Follow-up comments evaluated every 2-5 seconds
  • High-priority abort commands preempt within milliseconds
  • Streaming output for long-running tasks

Container Lifecycle

Key Facts (2025)

Usage Statistics

  • 115,000+ developers actively using Claude Code
  • 195 million lines of code processed weekly
  • 10x user growth since May 2025 launch
  • 90% of Claude Code itself written by AI

Revenue & Market

  • $500M+ annualized revenue from Claude Code
  • $130M estimated from Claude Code Web specifically
  • Part of Anthropic's $5B annualized revenue
  • 5.5x revenue jump with 300% user growth

Technical Specs

  • ~7GB universal container image
  • 4GB memory limit per sandbox
  • 4-hour JWT token expiry
  • 84% reduction in permission prompts with sandboxing

Availability

  • Pro plan ($20/month)
  • Max plans ($100-200/month)
  • Team & Enterprise users with premium seats

Use Cases

Security Considerations

Strengths

  1. Defense in Depth: Multiple isolation layers (gVisor + Seccomp + Network)
  2. Credential Protection: Real tokens never enter sandbox
  3. Auditable Operations: All git operations logged at proxy
  4. Ephemeral Storage: No data persistence after task completion
  5. Restricted Network: Allowlist-based egress filtering

Considerations

  1. Trust Model: Code executes on Anthropic infrastructure
  2. Network Access: Some external access required for packages
  3. Repository Access: OAuth grants broad repository permissions
  4. Rate Limits: Parallel tasks consume proportionally more

Comparison: Web vs CLI

FeatureClaude Code WebClaude Code CLI
ExecutionAnthropic cloud sandboxLocal machine
SecuritygVisor + proxy isolationOS-level sandboxing
GitHub AuthOAuth (scoped proxy)Personal tokens/SSH
Parallel TasksNative supportManual (multiple terminals)
NetworkAllowlist egress proxyFull network access
PersistenceEphemeral per-taskPersistent local
Mobile AccessYes (iOS app)No
OfflineNoYes

Architecture Principles

Sources

Technical research and documentation