Claude Code Base Action

The claude-code-base-action is a thin wrapper that installs and runs Claude Code. It does not enforce trust boundaries. Claude Code will read project-level configuration (like .claude/, CLAUDE.md, or .mcp.json) from the working directory.

Security Warning: The caller is responsible for ensuring the working directory and prompt are trusted. If your workflow processes untrusted input (e.g., from issues or fork pull requests), do not use this action. Instead, use anthropics/claude-code-action, which provides actor permission checks and protects project configuration.

You can trigger Claude Code by providing a direct string via the prompt input or by pointing to a text file via prompt_file. You must provide exactly one of these.

# Using a direct prompt
- name: Run Claude Code with direct prompt
  uses: anthropics/claude-code-base-action@beta
  with:
    prompt: "Your prompt here"
    allowed_tools: "Bash(git:*),View,GlobTool,GrepTool,BatchTool"
    anthropic_api_key: ${{ secrets.ANTHROPIC_API_KEY }}

# Or using a prompt from a file
- name: Run Claude Code with prompt file
  uses: anthropics/claude-code-base-action@beta
  with:
    prompt_file: "/path/to/prompt.txt"
    allowed_tools: "Bash(git:*),View,GlobTool,GrepTool,BatchTool"
    anthropic_api_key: ${{ secrets.ANTHROPIC_API_KEY }}

Model Context Protocol (MCP) configuration can be provided via a path to a JSON file (mcp_config) or as an inline JSON string. The configuration must follow the standard MCP format containing an mcpServers object.

# Using inline MCP config
- name: Run Claude Code with inline MCP config
  uses: anthropics/claude-code-base-action@beta
  with:
    prompt: "Your prompt here"
    mcp_config: |
      {
        "mcpServers": {
          "server-name": {
            "command": "node",
            "args": ["./server.js"],
            "env": {
              "API_KEY": "your-api-key"
            }
          }
        }
      }
    allowed_tools: "Bash(git:*),View,mcp__server-name__custom_tool"
    anthropic_api_key: ${{ secrets.ANTHROPIC_API_KEY }}

You can control the behavior of the Claude Code agent using the following inputs:

• system_prompt: Overrides the default system prompt.
• append_system_prompt: Appends instructions to the existing system prompt.
• max_turns: Limits the number of conversation turns to prevent infinite loops or excessive cost.

# Limit conversation to 5 turns
- name: Run Claude Code with limited turns
  uses: anthropics/claude-code-base-action@beta
  with:
    prompt: "Your prompt here"
    max_turns: "5"
    anthropic_api_key: ${{ secrets.ANTHROPIC_API_KEY }}

# Use a custom system prompt
- name: Run Claude Code with custom system prompt
  uses: anthropics/claude-code-base-action@beta
  with:
    prompt: "Build a REST API"
    system_prompt: "You are a senior backend engineer."
    anthropic_api_key: ${{ secrets.ANTHROPIC_API_KEY }}

Never hardcode API keys in your workflow files. Instead, use GitHub Actions secrets to manage sensitive credentials.

1. Add your key as a repository secret named ANTHROPIC_API_KEY in your GitHub repository settings.
2. Reference the secret in your workflow using the ${{ secrets.ANTHROPIC_API_KEY }} syntax.

# ✅ CORRECT - Uses GitHub secrets
anthropic_api_key: ${{ secrets.ANTHROPIC_API_KEY }}

Instead of using static API keys or OAuth tokens, you can use Workload Identity Federation to authenticate via GitHub OIDC. This requires the id-token: write permission on the job.

Note: Do not provide anthropic_api_key or claude_code_oauth_token when using federation, as static credentials take precedence and will prevent federation from working.

permissions:
  contents: read
  id-token: write

steps:
  - name: Run Claude Code with workload identity federation
    uses: anthropics/claude-code-base-action@beta
    with:
      prompt: "Your prompt here"
      anthropic_federation_rule_id: fdrl_xxxxxxxxxxxx
      anthropic_organization_id: 00000000-0000-0000-0000-000000000000
      anthropic_service_account_id: svac_xxxxxxxxxxxx

To provide Claude with environment-specific configuration (e.g., API URLs, database credentials, or feature flags) during its execution, use the claude_env input. This input accepts a YAML multiline string of key-value pairs. These variables are available to Claude at runtime.

- name: Deploy with custom environment
  uses: anthropics/claude-code-base-action@beta
  with:
    prompt: "Deploy the application to the staging environment"
    claude_env: |
      ENVIRONMENT: staging
      API_BASE_URL: https://api-staging.example.com
      DATABASE_URL: ${{ secrets.STAGING_DB_URL }}
      DEBUG: true
      LOG_LEVEL: debug
    allowed_tools: "Bash(git:*),View,GlobTool,GrepTool,BatchTool"
    anthropic_api_key: ${{ secrets.ANTHROPIC_API_KEY }}

You can specify the Node.js version used by the action by setting the NODE_VERSION environment variable. The default is 18.x.

- name: Run Claude Code with Node.js 20
  uses: anthropics/claude-code-base-action@beta
  env:
    NODE_VERSION: "20.x"
  with:
    prompt: "Your prompt here"
    anthropic_api_key: ${{ secrets.ANTHROPIC_API_KEY }}

You can provide Model Context Protocol (MCP) configurations using the --mcp-config flag within the claudeArgs string. The parser supports two modes:

1. Inline JSON: Provide a JSON string containing an mcpServers object. Multiple inline JSON configs will have their mcpServers merged.
2. File Path: Provide a path to a configuration file.

If you provide both inline JSON and a file path, the action attempts to merge the inline JSON objects. Note that if a file path is provided, the action's ability to merge it with inline JSON at parse-time is limited; ensure your file path includes all necessary servers if you are relying on a file-based config.

# Example of passing multiple MCP configs in claudeArgs
claudeArgs: '--mcp-config "{\"mcpServers\": {\"server1\": {\"command\": \"...\"}}}" --mcp-config /path/to/config.json'