Permissions from your roles
Tools are granted per role and time-boxed. An agent is capped by what its user may do.
AGENT SECURITY · MCP
Let agents act. Keep the keys.
A gateway between your AI agents and the tools they call. It checks every call against your roles, scans what goes in and comes out, asks a person when it matters, and keeps the receipt.
Approval neededWAITING
- agent
- support-bot
- for
- maria@acme.test
- tool
- stripe.refunds.create
- scan
- no secrets · no injection
What the gateway checks on every call
Not only which tool, but what flows through it, in both directions.
Secrets and injection scanning
Arguments and results are checked for leaked secrets and prompt injection, then blocked or flagged.
A person for the risky calls
Sensitive tools wait for a reviewer. The approval is single use and carries a written reason.
Tool changes flagged
If an MCP server adds or changes tools after you approved it, you are alerted.
Rate limits per tool
Quotas per grant, server and tool stop a runaway agent before it floods an API.
A receipt for every decision
Signed, portable proof of who allowed what, verifiable without trusting us.
One call, end to end
- 1 · AGENTCalls a tool
- 2 · ROLESAllowed for this user?
- 3 · SCANArguments clean?
- 4 · REVIEWApproved with a reason
- 5 · SCANResult clean?
- 6 · RECEIPTSigned and logged
A failed step stops the call and is logged with its reason. Built on MCP OAuth 2.0 and consent records.
Built on the Authorization engine
The roles your agents follow are the roles your users already have.