Skip to content
Documentation
Integrations
Anthropic

Anthropic

What it gives you — the investigation loop driven by native Claude, over the Messages API rather than an OpenAI-compatible shim — keyed against api.anthropic.com or a compatible in-cluster gateway.

Minimal config 

model:
  provider: anthropic
  model: claude-sonnet-5
  api_key_env: ANTHROPIC_API_KEY   # base_url defaults to https://api.anthropic.com

Pointed at an in-cluster gateway instead of the public API:

model:
  provider: anthropic
  base_url: http://claude-gateway.llm.svc:8080
  model: claude-sonnet-5
  # api_key_env omitted — keyless behind the gateway

Verify it locally 

kubectl -n runlore create secret generic runlore-secrets \
  --from-literal=ANTHROPIC_API_KEY='<key>'

Fire a test incident (see Alertmanager or hack/demo.sh) and confirm an investigation completes with a verdict:

kubectl -n runlore logs deploy/runlore | grep -E 'msg=findings|msg=incident'

Notes

  • base_url is optional — unset, it defaults to https://api.anthropic.com; set it to point at a compatible in-cluster gateway instead.
  • Requests carry anthropic-version: 2023-06-01.
  • Every forced-tool path RunLore relies on (the adversarial verify pass, the recall reranker, the eval judge, semantic KB validation, kb import --model) sends Anthropic’s native forced-tool_choice form ({"type":"tool","name":"…"}) — this is the reference-quality provider for those paths; see OpenAI-compatible for what happens on an endpoint that doesn’t honour it correctly.
  • effort (low/medium/high/max, sent as output_config.effort) and thinking: adaptive (sent as thinking: {type: "adaptive"}) are both supported, Anthropic-only. The client replays the model’s signed thinking blocks verbatim across the tool loop — give max_tokens headroom when thinking: adaptive is on, since thinking consumes output tokens. On the loop’s forced final-submission step, and for the always-forced verify pass, the client drops thinking and strips replayed thinking blocks for that one request (a forced tool_choice is incompatible with adaptive thinking in the API).
  • model.verify (a separate, optionally cheaper model for the adversarial verify pass) and model.embeddings (OpenAI-compatible /embeddings endpoint for hybrid recall — Anthropic has no embeddings API of its own) both work alongside this provider.
  • A cleartext API key is rejected on a public http:// host — use https:// or a loopback/in-cluster address.

Reference

Configuration → model for the full key reference.

SlackGitLab