Anthropic admin plane (usage, cost, compliance)

Claude Code telemetry tells you what runs on developer machines. The Anthropic admin plane tells you what the organization does: billed cost, per-workspace usage, org members and keys, the compliance activity feed. Four read-only sources cover it; this page wires the two central ones and summarizes their roster-side companions.

Source (kind)	What it reads	Credential
claude-api	usage & billed cost, model/workspace inventory, Claude Code analytics, API-side MCP/server-tool governance	Admin API key (admin_key)
claude-compliance	the compliance activity feed (evidence-grade events) + the org directory	activity-feed key + a distinct Compliance Access Key
claude-console	org IAM roster (members, roles) → SSO/SCIM posture findings	console credentials
claude-wif	non-human identities (service accounts svac_…, federated identities) + their permitted scope edges	WIF endpoint credentials

All are read-only and deny-closed: an empty credential means that feed is off and the product says so — never a fabricated empty inventory.

claude-api: cost, usage and API-side governance

{
  "sources": [{
    "name": "anthropic-org",
    "kind": "claude-api",
    "tenant": "<tenant-id>",
    "config": {
      "admin_key": "<admin-api-key-reference>",
      "cost_report": "true",
      "claude_code": "true"
    }
  }]
}

The keys that matter (from the shipped descriptor; defaults in parentheses):

• admin_key (secret) — the Anthropic Admin API key. Empty = offline catalog only.
• cost_report (true) — pull the billed cost report (daily, authoritative) alongside the derived usage estimate. The product keeps the two apart: estimates reconcile against billed figures, one source of cost per session, never both.
• lookback (24h) / cost_lookback (48h) / bucket_width (1d; also 1h, 1m) / max_pages — pull windows and pagination bounds.
• claude_code (false) — also pull the Claude Code Analytics feed (per-developer estimated cost by model) for chargeback.
• claude_code_shadow_auth (true) — with the analytics feed on, flag each developer whose Claude Code usage bills as customer_type=api — a personal/API key outside the org subscription, i.e. identity and spend riding an ungoverned key. Set false only if your org intentionally runs Claude Code on API billing.
• gateway (direct) — the deployment surface this org runs on (direct | claude-platform-aws | bedrock-mantle | bedrock-legacy | vertex | foundry). On a surface without the Admin API (Bedrock/Vertex/Foundry) the governance ingest degrades honestly with a posture finding instead of pretending an empty inventory.
• mcp_toolsets / server_tool_grants — operator-declared allow-sets for API-driven Claude agents (which MCP tools, which Anthropic server-tool types an agent may use). Each allowed entry becomes a permitted edge in module III, crossed against observed access — the same permitted-vs-observed diff as everywhere else. The agent_ref must be the agent’s external id as discovered at runtime, or the grant is an honest no-op rather than a false match.

The analytics feed has a named boundary

The Claude Code Analytics feed only tracks usage on the Claude API. Fleets on Claude Platform on AWS, Bedrock, Gemini Enterprise Agent Platform (formerly Vertex AI) or Microsoft Foundry are not in it — absence of findings there is not evidence of absence. For those surfaces the OTel plane is the observation you have.

claude-compliance: the evidence feed and the directory

{
  "sources": [{
    "name": "anthropic-compliance",
    "kind": "claude-compliance",
    "tenant": "<tenant-id>",
    "config": {
      "api_key": "<activity-feed-key-reference>",
      "compliance_access_key": "<compliance-access-key-reference>"
    }
  }]
}

Two distinct credentials, deliberately:

• api_key — an Admin API key with read:compliance_activities; pulls the activity feed (evidence-grade events).
• compliance_access_key — a separate key with read:compliance_org_data / read:compliance_user_data; enables the org directory ingest (orgs, users, roles, groups — including the SCIM-provisioning signal the Admin API cannot see). Empty = directory off, deny-closed.

The deletion scope (delete:compliance_user_data, used by the right-to- erasure path) is provisioned separately and dual-control gated — this read connector never holds it.

What you’ll see in the console

Billed and estimated spend, sliced by the dimensions the telemetry carries (team and project labels become first-class), in Cost & FinOps; org members, non-human identities and their scopes in Identity & NHI; posture findings (shadow auth, surface degradation, WIF footguns) in Security:

Honest limits

• Cost authority is the billed report. Usage-derived figures are estimates and are reconciled, never double-counted.
• The admin plane sees Anthropic-operated surfaces. Third-party-hosted Claude (Bedrock/Vertex/Foundry) is invisible to it — named explicitly via gateway, covered by the OTel plane.
• claude-console posture findings include a blind spot: the console cannot observe whether SSO/SCIM is enforced upstream — the finding says so rather than guessing.

Related

• Enterprise OTel for Claude Code — the per-session plane these org-level feeds complement.
• Budgets & FinOps guardrails — turn the cost stream into enforced limits.
• Connectors & coverage tiers — the full catalog.