Anthropic

Anthropic builds the Claude model family. AgentXchain connects to the Anthropic API via the api_proxy adapter for direct model access in governed turns.

Which adapter?

api_proxy with provider: "anthropic" — AgentXchain sends governed turn prompts to the Anthropic Messages API.

Prerequisites

• An Anthropic API key — get one from console.anthropic.com
• ANTHROPIC_API_KEY set in your environment
• agentxchain CLI installed

Configuration

{
  "runtimes": {
    "claude-qa": {
      "type": "api_proxy",
      "provider": "anthropic",
      "model": "claude-sonnet-4-6",
      "auth_env": "ANTHROPIC_API_KEY"
    }
  },
  "roles": {
    "qa": {
      "runtime": "claude-qa",
      "mandate": "Review code, find bugs, verify correctness",
      "authority": "review_only"
    }
  }
}

Available models

Model	ID	Input (per 1M)	Output (per 1M)	Best for
Claude Opus 4.6	claude-opus-4-6	$5.00	$25.00	Complex architecture, deep reasoning
Claude Sonnet 4.6	claude-sonnet-4-6	$3.00	$15.00	Balanced coding + review (recommended default)
Claude Haiku 4.5	claude-haiku-4-5-20251001	$1.00	$5.00	Fast review, linting, high-volume QA

Multi-role example

Use different Claude models for different roles to optimize cost and quality:

{
  "runtimes": {
    "opus-architect": {
      "type": "api_proxy",
      "provider": "anthropic",
      "model": "claude-opus-4-6",
      "auth_env": "ANTHROPIC_API_KEY"
    },
    "sonnet-dev": {
      "type": "api_proxy",
      "provider": "anthropic",
      "model": "claude-sonnet-4-6",
      "auth_env": "ANTHROPIC_API_KEY"
    },
    "haiku-qa": {
      "type": "api_proxy",
      "provider": "anthropic",
      "model": "claude-haiku-4-5-20251001",
      "auth_env": "ANTHROPIC_API_KEY"
    }
  },
  "roles": {
    "architect": { "runtime": "opus-architect", "mandate": "Design system architecture", "authority": "review_only" },
    "dev": { "runtime": "sonnet-dev", "mandate": "Implement features", "authority": "proposed" },
    "qa": { "runtime": "haiku-qa", "mandate": "Review and test", "authority": "review_only" }
  }
}

Verify the connection

export ANTHROPIC_API_KEY="sk-ant-..."
agentxchain connector check

The connector probe sends a test request to verify your API key and model access.

Budget configuration

AgentXchain has bundled cost rates for all Claude models. Budget enforcement works automatically:

{
  "budget": {
    "max_cost_per_run": 5.00,
    "warn_cost_per_run": 3.00
  }
}

Gotchas

• 200K context window: Claude models support large contexts, which is helpful for large dispatch bundles. AgentXchain automatically manages prompt construction within token limits.
• Rate limits: Anthropic applies per-account rate limits. For high-volume governed runs with many turns, consider requesting higher limits via the Anthropic console.
• Streaming: AgentXchain uses single-request (non-streaming) API calls. The full response is returned at once.