Every AI surface in SignalGuard runs on Anthropic's Claude models. Connecting your own Anthropic API key moves that inference onto your account: your contract, your rate limits, your data-retention settings, and a usage record you can audit in your own console.
What your Claude key does in SignalGuard
Claude is not a data source in SignalGuard — it is the reasoning layer on top of the 70+ collected signals. It writes the AI threat brief for a scan, it scores collected chatter for severity, and it runs Signal Scout — the agent-directed open-web sweep that goes looking for what a fixed source list would miss for a specific event or artist.
Scout is the surface that reads your key today. A run uses Claude's web search to sweep the open web against the focus an operator set, retrieval on Claude Sonnet 5 to pull the candidate material, and Claude Haiku 4.5 to classify each finding. What lands is cited evidence, never a score: Scout has no entry in the scored signal roster, so a finding cannot move a posture number on its own. It gets filed with its source, its verification state, and its finding class, for an operator to read.
The rest of the AI layer runs on a shared SignalGuard Anthropic account by default, so nothing is dark before you connect a key. What a customer key changes is whose contract the inference sits on — which, for security teams with their own model-governance and retention requirements, is usually the whole point.
How the integration works
Paste your API key into the Anthropic tile at /integrations. SignalGuard validates it with a single GET /v1/models call — the cheapest probe that proves a key works, and one that spends no tokens; a messages probe would bill you to find out whether your own key is valid. The key is then stored AES-256-GCM encrypted at rest, decrypted only in server memory at request time, and never sent to the browser or written to logs. Every Claude call is made server-side.
Deliberately, the validator does not check for an sk-ant- prefix. That prefix is a convention, not a contract, and a validator that hard-codes it starts rejecting valid keys the day the format changes.
Common use cases
- Running event-security AI on your own Anthropic contract for procurement or vendor-risk sign-off
- Holding model inference under your organisation's own data-retention configuration
- Auditing exactly what SignalGuard asked a model, from your own Anthropic console
- Sizing Signal Scout usage against a rate limit you control rather than a shared ceiling
- Charging AI spend to the security programme's own cloud budget line
- Enterprise deployments where a shared vendor model account is not an acceptable posture
What you need
An Anthropic account with API access and a funded balance. Create a key at console.anthropic.com → Settings → API keys. Billing is usage-based and goes to Anthropic directly — SignalGuard does not resell Claude capacity and takes no margin on it. As a reference point, one Signal Scout run costs roughly $0.21 at current rates.
FAQ
Do I need an Anthropic subscription? No. SignalGuard's AI surfaces work out of the box on a shared account. Connect your own key when you want the spend, the ceiling and the retention terms on your contract instead of ours.
Where do I add my key? In your SignalGuard workspace at /integrations. Open the Anthropic (Claude) tile and paste the key.
Is my key secure? Yes. It is AES-256-GCM encrypted at rest, decrypted only at request time on the server, and never logged or sent to the browser. Revoke or swap it from the same tile at any time.
Connect Anthropic
Connect Claude in your SignalGuard workspace → and read the BYOK explainer for how SignalGuard stores and resolves customer keys.