User identity
OpenID ConnectOIDC carries authenticated user identity from an existing identity provider. Arcane keeps that user visible when an agent acts on their behalf.
arcane / product
The AI agent control plane.
Arcane gives each agent a verified identity, evaluates what it may do at the action boundary, and preserves the evidence behind every allow, deny, or escalate decision—including what the task cost.
- Category
- AI agent control plane
- Identity
- User · agent · workload
- Decision
- Allow · deny · escalate
- Boundaries
- SDK · gateway · MCP
- Analysis
- Tokens · task · spend
Need the vocabulary? Read the AI agent control plane glossary.
1.0 · Primitives
What does an AI agent control plane do?
An AI agent control plane discovers acting agents, verifies their identity, evaluates and enforces permissions, understands their tasks, attributes token spend, and records evidence. Arcane keeps those functions connected as agents act across tools and systems.
2.0 · Standards
Which standards does Arcane build on?
Arcane composes established identity and authorization standards instead of replacing them. The control plane adds the agent, task, policy, and evidence context those protocols do not define on their own.
Delegation
OAuth 2.0 Token ExchangeRFC 8693 defines token exchange and actor context for delegation. Arcane uses the same model to preserve who delegated authority and which agent received it.
Tool transport
Model Context ProtocolMCP defines how clients reach tools and authorize HTTP transports. Arcane adds identity-aware policy and decision evidence at the tool-action boundary.
3.0 · Real examples
Why is service-account IAM insufficient for AI agents?
A service account identifies a workload but usually loses the delegating user, task, and per-action context. Arcane keeps those entities together so policy and audit can distinguish one agent action from another.
Scenario
Service-account IAMwithout Arcane
Composite identitywith Arcane
Agent leaks a credential
Service-account IAM
Static key with broad scope. Rotation requires a deploy. Blast radius = every system the key touches.
With Arcane
Credentials are task-bound with a short TTL — minutes, not days. The leaked credential expires before anyone notices.
Agent makes an unexpected call
Service-account IAM
No per-call enforcement. Allowed by default. You find out from the audit log — after the fact.
With Arcane
Policy evaluates the call at the boundary. Out-of-scope tools are denied. The decision and reason are surfaced.
Auditor asks who did what
Service-account IAM
Service-account-X did everything. Attribution to a specific user or task is reconstructed from logs.
With Arcane
Each request carries user · agent · workload · task. Attribution is a query, not a forensics project.
4.0 · The frame
How does Arcane reduce AI agent risk?
Arcane replaces standing access with short-lived grants, evaluates policy before each action, and keeps operator approval explicit. Security teams can expand agent use without granting every workflow broad, permanent credentials.
Approval lanes — explicit
Escalations become review points instead of hidden side-effects inside an agent run.
Static keys — gone
Short-lived, task-bound tokens replace standing credentials. The blast radius shrinks with the TTL.
Policy — proposed, you approve
Arcane drafts. A human acks, edits, or rejects. Security keeps the pen.
Inline
Per-call evaluation
Signed
Every decision
Task-bound
Every token
None
Standing credentials