claude-code-action/docs/experimental.md

Experimental Features

Note: Experimental features are considered unstable and not supported for production use. They may change or be removed at any time.

Automatic Mode Detection

The action intelligently detects the appropriate execution mode based on your workflow context, eliminating the need for manual mode configuration.

Interactive Mode (Tag Mode)

Activated when Claude detects @mentions, issue assignments, or labels—without an explicit prompt.

• Triggers: @claude mentions in comments, issue assignment to claude user, label application
• Features: Creates tracking comments with progress checkboxes, full implementation capabilities
• Use case: Interactive code assistance, Q&A, and implementation requests

- uses: anthropics/claude-code-action@v1
  with:
    anthropic_api_key: ${{ secrets.ANTHROPIC_API_KEY }}
    # No prompt needed - responds to @claude mentions

Automation Mode (Agent Mode)

Automatically activated when you provide a prompt input.

• Triggers: Any GitHub event when prompt input is provided
• Features: Direct execution without requiring @claude mentions, streamlined for automation
• Use case: Automated PR reviews, scheduled tasks, workflow automation

- uses: anthropics/claude-code-action@v1
  with:
    anthropic_api_key: ${{ secrets.ANTHROPIC_API_KEY }}
    prompt: |
      Check for outdated dependencies and create an issue if any are found.
    # Automatically runs in agent mode when prompt is provided

How It Works

The action uses this logic to determine the mode:

1. If prompt is provided → Runs in agent mode for automation
2. If no prompt but @claude is mentioned → Runs in tag mode for interaction
3. If neither → No action is taken

This automatic detection ensures your workflows are simpler and more intuitive, without needing to understand or configure different modes.

Advanced Mode Control

For specialized use cases, you can fine-tune behavior using claude_args:

- uses: anthropics/claude-code-action@v1
  with:
    prompt: "Review this PR"
    claude_args: |
      --max-turns 20
      --system-prompt "You are a code review specialist"
    anthropic_api_key: ${{ secrets.ANTHROPIC_API_KEY }}

Network Restrictions

For enhanced security, you can restrict Claude's network access to specific domains only. This feature is particularly useful for:

• Enterprise environments with strict security policies
• Preventing access to external services
• Limiting Claude to only your internal APIs and services

When experimental_allowed_domains is set, Claude can only access the domains you explicitly list. You'll need to include the appropriate provider domains based on your authentication method.

Provider-Specific Examples

If using Anthropic API or subscription

- uses: anthropics/claude-code-action@v1
  with:
    anthropic_api_key: ${{ secrets.ANTHROPIC_API_KEY }}
    # Or: claude_code_oauth_token: ${{ secrets.CLAUDE_CODE_OAUTH_TOKEN }}
    experimental_allowed_domains: |
      .anthropic.com

If using AWS Bedrock

- uses: anthropics/claude-code-action@v1
  with:
    use_bedrock: "true"
    experimental_allowed_domains: |
      bedrock.*.amazonaws.com
      bedrock-runtime.*.amazonaws.com

If using Google Vertex AI

- uses: anthropics/claude-code-action@v1
  with:
    use_vertex: "true"
    experimental_allowed_domains: |
      *.googleapis.com
      vertexai.googleapis.com

Common GitHub Domains

In addition to your provider domains, you may need to include GitHub-related domains. For GitHub.com users, common domains include:

- uses: anthropics/claude-code-action@v1
  with:
    anthropic_api_key: ${{ secrets.ANTHROPIC_API_KEY }}
    experimental_allowed_domains: |
      .anthropic.com  # For Anthropic API
      .github.com
      .githubusercontent.com
      ghcr.io
      .blob.core.windows.net

For GitHub Enterprise users, replace the GitHub.com domains above with your enterprise domains (e.g., .github.company.com, packages.company.com, etc.).

To determine which domains your workflow needs, you can temporarily run without restrictions and monitor the network requests, or check your GitHub Enterprise configuration for the specific services you use.