AI agents need real accounts to do useful work: read a calendar, post to Slack, push to GitHub. Giving an agent a long-lived credential puts that credential within reach of whatever the agent reads and runs.
Cred sits in between. By default, provider tokens stay in an encrypted vault on your server, and agents get short-lived Delegation Tokens bound to their own keys. Cred checks a fresh proof on every request, makes the call and records it in an audit log.
Cred is open source under Apache-2.0, with TypeScript and Python SDKs, an MCP server and framework adapters. You run it on your own infrastructure, alongside your identity and security tools.
The design is published as an individual Internet-Draft.
Your identity provider says who a person is. Cred handles what their agent can reach on their behalf.
Secrets managers hold infrastructure secrets you own. Cred holds the OAuth tokens people connect for their agents to use.
Want to talk? Email kieran@kierans.net.