Anthropic model provider

Claude Code compatibility

The Anthropic model provider is fully compatible with Claude Code, which allows you to run Claude Code directly within your Buildkite pipelines, enabling automated code generation, refactoring, and testing in your CI/CD environment.

Supported models

Buildkite supports all current Anthropic Claude models, including Claude Sonnet 4.6, Claude Sonnet 4.5, Opus 4.1, and Haiku 4.5.

Using Claude Code in pipelines

Claude Code's headless mode (claude -p "prompt") lets you run Claude as a non-interactive step in Buildkite Pipelines. To connect Claude Code to the Buildkite model provider, set the following environment variables in your pipeline step:

env:
  ANTHROPIC_BASE_URL: "$BUILDKITE_AGENT_ENDPOINT/ai/anthropic"
  ANTHROPIC_API_KEY: "$BUILDKITE_AGENT_ACCESS_TOKEN"

A basic pipeline example:

steps:
  - label: ":claude: Code review"
    command: |
      claude -p "Review the changes in this PR and suggest improvements" \
        --permission-mode bypassPermissions
    env:
      ANTHROPIC_BASE_URL: "$BUILDKITE_AGENT_ENDPOINT/ai/anthropic"
      ANTHROPIC_API_KEY: "$BUILDKITE_AGENT_ACCESS_TOKEN"

The --permission-mode bypassPermissions flag is required for CI environments where there is no human to approve tool use prompts.

Running as a non-root user

Claude Code refuses to run with --permission-mode bypassPermissions as the root user for security reasons. If your Buildkite agent runs as root, use su to switch to a non-root user:

steps:
  - label: ":claude: Analyze failures"
    command: |
      su buildkite -c 'HOME=/home/buildkite claude -p "Analyze the build failures" --permission-mode bypassPermissions'
    env:
      ANTHROPIC_BASE_URL: "$BUILDKITE_AGENT_ENDPOINT/ai/anthropic"
      ANTHROPIC_API_KEY: "$BUILDKITE_AGENT_ACCESS_TOKEN"

Base URL

Once you have connected your Buildkite organization to your Anthropic model provider, you can access your Anthropic Claude models through the Claude API, by appending these endpoints to the relevant Buildkite model provider API endpoint as the base URL:

https://agent.buildkite.com/v3/ai/anthropic

Supported endpoints

The following Claude API endpoints are available through Buildkite model provider API:

• POST /v1/messages endpoint: Generates completions and chat responses. Token usage is automatically tracked for billing.
• POST /v1/messages/count_tokens endpoint: Calculates token usage before making requests to optimize costs.
• GET /v1/models endpoint: Retrieves all available Anthropic models.
• GET /v1/models/{model_id} endpoint: Gets information about a specific model's capabilities and limits.

These endpoints are accessed by appending them to the end of your Buildkite model provider API's base URL—for example, to access the Claude API POST /v1/messages endpoint from your Buildkite agent, use the following URL:

https://agent.buildkite.com/v3/ai/anthropic/v1/messages

Authentication methods

The Anthropic model provider supports two authentication header formats, both of which use a job token for authentication.

Authorization header (standard Agent API)

-H "Authorization: Token $BUILDKITE_AGENT_ACCESS_TOKEN"

x-api-key header (Claude SDK compatible)

-H "x-api-key: $BUILDKITE_AGENT_ACCESS_TOKEN"

Basic example

Here's a simple pipeline that generates unit tests for your code:

steps:
  - label: "Failure analysis"
    command: |
      curl -X POST "$BUILDKITE_AGENT_ENDPOINT/ai/anthropic/v1/messages" \
        -H "Content-Type: application/json" \
        -H "x-api-key: $BUILDKITE_AGENT_ACCESS_TOKEN" \
        -d '{
          "model": "claude-sonnet-4-5",
          "max_tokens": 1000,
          "messages": [
            {
              "role": "system",
              "content": "..."
            },
            {
              "role": "user",
              "content": "Analyze the test failures in this log"
            }
          ]
        }'

Rate limits

The following rate limits apply to Anthropic API requests:

Request rate limiting

• Default limit: 50 requests per minute

Input token rate limiting

• Default limit: 50,000 input tokens per minute per provider.

• Token calculation: total_input_token = cache_creation_input_tokens + input_tokens.

To request a higher rate limit for your Buildkite organization, please contact support@buildkite.com.

Response formats

Anthropic provider supports both:

• Non-streaming responses: Complete responses returned after processing.

• Streaming responses: Real-time response chunks for long-running completions.