Sanad GitHub

Short-lived credentials for AI agents.

Sanad is an open-source gateway in front of your MCP servers. Every agent request is verified back to an accountable human, granted a passport that expires in minutes, and written to a tamper-evident log. When something goes wrong, you can kill all of it in under a minute.

View on GitHub $ npm i @getsanad/sdk

Passports, not standing keys

Each request earns a credential scoped to one server and one task, expiring in seconds to minutes. The upstream server never sees your real keys.

Delegation can only narrow

Signed chains from human to agent to sub-agent. Every hop can hand over less authority than it holds, never more. The gateway rejects anything else.

A kill switch that means it

Revoke a principal and everything rooted in it loses access within a minute. Every decision lands in a tamper-evident audit log.

In classical scholarship, a sanad is the chain of transmission that makes a claim trustworthy: who heard it from whom, back to the source. Sanad builds that chain for AI agents — from the accountable human, through every delegation, to the action taken.