Overview
PolicyLayer is the control plane for your MCP fleet.
Your agents connect to PolicyLayer instead of connecting directly to upstream MCP servers. PolicyLayer authenticates each caller, evaluates the requested tool call against policy, and only forwards the request when the policy allows it.
That gives you one enforcement point across MCP servers you own, MCP servers you deploy, and MCP servers you only consume.
What PolicyLayer controls
- Servers. The upstream MCP servers your agents need to reach, such as Stripe, GitHub, Linear, or internal tools exposed over HTTP.
- Grants. Labelled bearer tokens for MCP clients and automations such as
alice-laptop,ci-runner, orsupport-agent. - Policies. Deterministic rules that allow, deny, hide, or rate-limit tool calls.
- Roles. Org roles that separate policy authors, credential admins, and read-only viewers.
What you get
- Hard enforcement before execution. Policies are evaluated by the proxy before the upstream server receives the call.
- Scoped grants. Each grant belongs to one server and can carry its own policy. Rotate or revoke one grant without breaking the others.
- Credential separation. Agents use PolicyLayer grants. Upstream OAuth tokens and static headers stay server-side, encrypted at rest, and are not returned to MCP clients.
- Grant token handling. Grant tokens are hashed for lookup and encrypted in storage.
- Redacted decision logs. Proxy logs show the grant, server, tool, decision, policy, argument keys, status, and timestamp without storing argument values.
- Live policy changes. Policy edits apply to the next request. Clients do not need to reconnect.
Start with the Quick start to set up your first server, policy, and grant.