MCP security gateway
MCP security gateway for AI coding agents
MCP makes tools easy to connect to agents. That also makes permissions harder to reason about when each server can reach APIs, credentials, files, shell commands, or internal systems.
Agent Access Broker is designed as a local MCP gateway: the agent calls brokered tools, the gateway maps the request to a Grant Lease, and the connector uses the real credential only when policy allows it.
This is not a full enterprise IAM replacement. It is a narrow runtime firewall for vibe coding agents that need to do real work without seeing raw secrets.
Route MCP tool calls through a local policy engine.
Map temporary tokens to Grant Leases, not raw credentials.
Attach TTL, resource, service, action, and approval constraints.
Emit receipts that help teams audit AI agent actions.
Where MCP permission risk appears
An MCP server may hide broad access behind a friendly tool name. A single tool can reach files, APIs, cloud CLIs, databases, deploy systems, or payment data unless the gateway narrows the action and resource boundary.
Brokered MCP gateway pattern
Put a local policy layer between the agent and each MCP connector. The gateway checks the grant lease, calls the connector only for allowed actions, and emits receipts that show what happened.
Common questions
Does MCP replace credential permissions?
No. MCP connects tools to agents, but you still need policy, scope, TTL, approval, and audit controls around the credentials those tools can use.
What should an MCP security gateway log?
Log the agent, grant, connector, action, resource, decision, timestamp, and whether the request was allowed, denied, approved, or revoked.