FAQ
An AI agent control plane is the layer that establishes agent identity, evaluates permissions, enforces policy, and records decisions across tools and systems. Arcane applies those controls to each meaningful agent action.
Arcane binds three verified entities: the user or system that delegated the work, the registered agent that acts, and evidence about the workload where it runs. Policy evaluates that composite identity instead of trusting an agent's self-description.
No. Arcane uses trusted identity and workload issuers as evidence inputs rather than replacing them. Current workload-verification modes include generic OIDC JWTs, Kubernetes projected service-account tokens, GCP identity tokens, and SPIFFE JWT-SVIDs.
No. Credential handling is one part of the control plane. Arcane mints short-lived identity tokens, evaluates the requested action, and then proxies execution or vends scoped downstream access according to the resource's enforcement mode.
Two paths. The SDK wraps tool invocations inside your agent process. The transparent proxy sits in front of any HTTP service — MCP servers, internal APIs, third-party SaaS — and evaluates the same policy.
A structured record connects the timestamp, authority, agent, workload, task, resource, matched policy, outcome, enforcement result, and reason. Tamper-evident audit and formal retention guarantees remain roadmap controls rather than shipped claims.
Arcane Cost Center attributes token use and spend to the agent, task, model, team, and delegating authority behind it. Teams can compare equivalent work and optimize token use while keeping task outcome and quality in view.
Arcane is designed for self-managed deployment, with enforcement available through SDK, proxy, MCP-aware, and ext_authz boundaries. The exact network and trust boundary depends on the deployment topology and should be verified during implementation.
We're working directly with a small number of teams whose agents touch real data — internal or customer-facing. Talk to founders to start a conversation.