v1.4.0

Using Noir with Anthropic

Use Claude models with Noir through Anthropic's native Messages API.

The anthropic preset talks to Anthropic's native Messages API (/v1/messages), not its OpenAI-compatible layer. The endpoint schema Noir asks for is sent as output_config.format, so Claude's reply is constrained to it.

Setup

  1. API Key: Create one in the Claude Console.
  2. Model: Pick a model ID such as claude-opus-5-5 or claude-sonnet-5-5.

Usage

noir scan ./myapp \
     --ai-provider=anthropic \
     --ai-model=claude-opus-5-5 \
     --ai-key=sk-ant-...

Without --ai-key, Noir reads NOIR_AI_KEY, then ANTHROPIC_API_KEY:

export ANTHROPIC_API_KEY=sk-ant-...
noir scan ./myapp --ai-provider=anthropic --ai-model=claude-sonnet-5-5

Notes

  • A gateway in front of the Messages API works too: --ai-provider=https://gw.example/anthropic/v1/messages. Any api.anthropic.com URL uses the native API unless it ends in /chat/completions.
  • If a model or gateway rejects the structured-output request, Noir drops it for the rest of the run and relies on the prompt's JSON instructions.
  • Models with fixed sampling (Claude Opus 4.7 and later, and Claude 5 models) are sent no temperature. If another model rejects it, Noir drops it for the rest of the run.
  • Each reply is capped at 16,000 output tokens (4,096 for Claude 3, 8,192 for Claude 3.5). A reply cut at that cap is reported as truncated, and the complete endpoints before the cut are kept.
  • --ai-agent uses Claude's native tool calling (anthropic is in the default --ai-native-tools-allowlist).
  • --ai-stream does not apply to this provider.