Configure Anthropic Claude as Cynative's LLM Provider

Set up Cynative with Anthropic’s Claude models. Configure API key, model selection, reasoning effort, and prompt caching for cost efficiency.

Cynative supports Anthropic’s Claude family through the Bifrost anthropic provider. Claude models are well-suited to security research workloads: large context windows, strong instruction-following, and native support for extended thinking (reasoning). Cynative also automatically enables prompt caching for Anthropic, so repeated tool schemas and system prompts are billed at the cached read rate.

Quick start

export ANTHROPIC_API_KEY=sk-ant-...
export CYNATIVE_LLM_PROVIDER=anthropic
export CYNATIVE_LLM_MODEL=claude-opus-4-7
cynative -p "enumerate S3 buckets and check each one for public access"

No ~/.cynative/config.yaml file is required.

YAML configuration

llm:
  provider: anthropic
  model: claude-opus-4-7
  api_key: env.ANTHROPIC_API_KEY    # optional — falls back to ANTHROPIC_API_KEY env var
  network_config:
    default_request_timeout_in_seconds: 300
    max_retries: 3
    extra_headers:                  # optional; must be set in YAML (map — no env form)
      anthropic-beta: tools-2024-04-04

Authentication

Get an API key from console.anthropic.com/settings/keys. Cynative resolves the key in this order:

1. llm.api_key in YAML (or CYNATIVE_LLM_API_KEY env var).
2. llm.keys[] in YAML (for multi-key load balancing or model filters).
3. ANTHROPIC_API_KEY — the canonical fallback.

Setting both llm.api_key and llm.keys[] is rejected at startup.

Reasoning (extended thinking)

Claude models that support extended thinking expose it through Cynative’s provider-portable reasoning_effort and reasoning_max_tokens fields:

llm:
  provider: anthropic
  model: claude-opus-4-7
  api_key: env.ANTHROPIC_API_KEY
  reasoning_effort: high            # pick ONE of the two forms:
  # reasoning_max_tokens: 8192      # explicit budget (takes priority over effort)

Or via environment variables:

export CYNATIVE_LLM_REASONING_EFFORT=high

How reasoning maps to Anthropic’s API

Cynative field	Behavior
reasoning_effort	Derives a thinking budget from the effort level. On Claude Opus 4.7+ (adaptive thinking), any non-none effort simply switches on adaptive thinking.
reasoning_max_tokens	Sets an explicit thinking budget. On pre-4.7 models, the minimum is 1,024 — smaller budgets are rejected at request time. On Opus 4.7+, any numeric value switches on adaptive thinking; the exact number is ignored.
Both set together	reasoning_max_tokens takes priority and the effort level is ignored.
reasoning_effort: none	Explicitly disables thinking. Combining none with reasoning_max_tokens is rejected at config validation.

Prompt caching

Cynative marks the stable prefix of every request with ephemeral cache breakpoints — the tool schemas, the system prompt, and the last two conversation turns. Anthropic’s API caches that prefix and bills cached reads at a significant discount.

Key points:

• Cache TTL: Anthropic’s default five-minute ephemeral TTL. Back-to-back calls (e.g. unattended --auto-approve runs) benefit the most; long human-approval pauses may let the cache expire.
• Nothing to configure. Caching is on automatically for the anthropic provider (and for Claude models on bedrock and vertex).
• Write vs. read cost. The first call in a window pays a write cost; subsequent calls within the window pay only the (cheaper) read cost. The Anthropic usage dashboard shows the split.

Environment variable reference

Environment variable	YAML equivalent	Description
ANTHROPIC_API_KEY	llm.api_key	Anthropic API key (sk-ant-...).
CYNATIVE_LLM_PROVIDER	llm.provider	Must be anthropic.
CYNATIVE_LLM_MODEL	llm.model	Model id (e.g. claude-opus-4-7, claude-sonnet-4-5).
CYNATIVE_LLM_API_KEY	llm.api_key	Generic key override (takes priority over ANTHROPIC_API_KEY).
CYNATIVE_LLM_NETWORK_CONFIG_DEFAULT_REQUEST_TIMEOUT_IN_SECONDS	llm.network_config.default_request_timeout_in_seconds	Per-request timeout in seconds. Cynative default: 300.
CYNATIVE_LLM_NETWORK_CONFIG_MAX_RETRIES	llm.network_config.max_retries	Retry attempts on 429/5xx errors. Cynative default: 3.
CYNATIVE_LLM_REASONING_EFFORT	llm.reasoning_effort	none, minimal, low, medium, or high.
CYNATIVE_LLM_REASONING_MAX_TOKENS	llm.reasoning_max_tokens	Explicit thinking budget in tokens (integer, minimum 1024 for pre-4.7 models).

Links

• Anthropic API reference
• Bifrost Anthropic provider source