welikecode/mintkey

Mintkey is a self-hostable credential broker built specifically for AI agents, programs that act on their own to call outside services. Normally, giving an agent access to a service means handing it the real password or API key directly, which means the agent holds a secret it could leak or misuse. Mintkey is built to avoid that. You register your services and their real credentials with Mintkey once. When an agent needs to call one of those services, it asks Mintkey for a short-lived, scoped token that only works for that specific service and expires after about ten minutes. The agent sends its requests through Mintkey's egress proxy, which swaps in the real credential only at the moment the request reaches the destination service. At no point does the agent itself hold or see the underlying secret. Every request that passes through the system is recorded in a per-request audit trail using a mandatory hash chain, so an operator can review exactly what each agent accessed and when. Mintkey also runs as an MCP server, a standard interface that lets agents discover which services are available and request tokens for them without any manual, out-of-band setup. The project makes clear what it is not: it is not a general secrets manager meant for humans or CI and CD pipelines, not an agent runtime, not an inbound API gateway, and not a tool for preventing prompt injection, though it does limit what a compromised agent can reach. It can be run locally with a single docker compose command, and a basic setup is expected to become healthy within about two minutes. Mintkey is currently pre-alpha, its interface is marked experimental, and the README explicitly says not to use it in production yet. It is licensed under Apache-2.0. The full README is longer than what was shown.