MCP gateway or direct connections - Oriant

When someone connects Claude Code or Cursor straight to an MCP server, the AI acts in that system as them. That works. What’s missing is everything the company needs around it: a record, a say in which tools are used, and a way to stop it. A gateway sits in the middle. The AI client connects to the gateway, and the gateway connects to the server.

Side by side

	Direct	Through Oriant
Who the call acted for	Only the server’s own logs, if any	Every call, with the member and AI client
Which tools people may use	Whatever their account allows	Decided per team by MCP Policies
Limits on arguments	None	Argument rules, such as one Linear team
Server tokens	Stored in each AI client	Stored encrypted by Oriant, never handed to the AI client
Removing access	Revoke in every server, per person	Remove the connection, or change the policy
If the gateway is down	Not affected	Calls stop. A call that can’t be recorded isn’t sent

When direct is fine

One person, a personal account, no company data. Then the server’s own permissions are enough.

When a gateway makes sense

People use AI on company systems: tickets, customer data, cloud accounts. You need to answer “what did AI do, and for whom?”, and you want read and write to differ from team to team.

What a gateway can’t do

It only sees calls that go through it. Someone can still add a server straight to their own config. The endpoint agent finds those on each machine, so you know who to move over.

Last modified on November 9, 2026