Sign in
Analytics
Spend, enforcement, and leaderboards across the fleet.
Spend per day
Blocks & flags per day
Top spenders
Top agents
Models
Rule & detector hits
a rule firing constantly may be too broadAgents
Governance readiness
Agents
last 30 days · auto-discovered on first callMy Governance
Security posture
Security tools & enforcement
applies to your company only · leave on Inherit to use the account defaultSecurity providers
prompt-injection & content-safety detectors · add your own key per provider · stored encryptedProvider API keys
your own LLM provider keys · stored encrypted · override the account default for your companyContent rules
your company's own rules — added on top of the account defaultsEnd-customers
End-customers
the customers your agents serve · last 30 days · auto-discovered on first call · setcustomer_id in setContextCustomers
How to onboard a new customer
acmecorp.Tenants
Early access waitlist
Settings
Tenant knobs
Policies
Content rules
Groups
Users
Security tools
Third-party detectors run inline on every LLM call. Toggle a tool on or off without losing your API key.
HIPAA PHI Redaction Premium
Redacts 18 HIPAA Safe Harbor identifiers before they reach the LLM
Lakera Guard
Prompt injection and jailbreak detection
Approved endpoints
Pin each provider to the destination its calls must use. Any agent calling that provider is restricted to the endpoint set here. Leave a provider blank to allow it anywhere.
Model pricing
Per-model rates in USD per 1,000,000 tokens. Overrides layer over the built-in defaults (shown as placeholders) and resolve to a model id by longest prefix (e.g. claude-sonnet-4 covers claude-sonnet-4-6). Leave a field blank to keep the default; blank rows are ignored.
MCP & tool-call governance
Govern the actions your agents take through MCP servers and tools — the layer where a prompt-injection turns into data loss. Enforced in-process before the tool call runs.
Approved MCP servers
Allow-list of server endpoints a tool call may reach. Empty = unrestricted. Patterns: *, *.mcp.acme.com, http://localhost:*, or a bare host. Any call to a server not on this list is blocked before it leaves the process.
Blocked tools
Tool-name glob patterns to hard-block (e.g. delete_*, execute_shell). One per line.
Lethal-trifecta guard
Blocks an exfiltration-capable tool call once a session has also touched private data and untrusted content — the combination a prompt-injection attack needs to steal data (the lethal trifecta). Classification uses built-in heuristics unless overridden.
Provider keys (credential vault)
Store each LLM provider's API key once, centrally. SDK processes receive the right key through the policy bundle — developers never put provider keys in env files or code, and rotating a key here updates every agent within one bundle refresh.
Scope Account default applies to everyone; a key scoped to an end-customer overrides the default for that customer's traffic only. Keys are AES-256-GCM encrypted at rest and can never be read back from this page.
CREDENTIAL_ENC_KEY (32-byte base64/hex) on the backend to enable it.