Thoughts on building DeepSecret
For years I kept seeing the same pattern, and I was part of it too.
Production credentials in Slack threads. Kubeconfigs shared in DMs. Someone pastes an rclone command with creds embedded into an 80-person channel because there's an incident, three engineers are debugging on a Google Meet, and the logs need to move now. Someone uploads a .env to unblock a teammate. Someone else drops a temporary token into a thread that will live in search forever.
It wasn't that people were irresponsible. It was that workflow convenience always wins under pressure.
After enough of these moments I started sketching what I'd actually want from a system I used daily with infra and security teams.
On the technical side: keys generated locally. Encryption and decryption client-side. I'd own the ciphertext, in my own S3-compatible bucket, so storage size was never a limit. Keys should be rotatable, and storage should be movable to another provider, without turning the whole thing into a migration project. And it had to be a CLI; I live in the terminal.
On the operational side: easy teammate sharing, sending to groups without ceremony, exchanging encrypted data naturally during real engineering work, not fighting the tool every time something urgent happens. And the one that kept biting was: moving files between machines in private subnets, without bastion gymnastics or one-off bucket policies, when all I wanted was to focus on the actual incident.
That's what DeepSecret is.
It's a CLI that generates keys and does end-to-end encryption between teammates. The backend stores metadata, but never plaintext content. Encryption happens on your machine. Ciphertext lives in your own S3-compatible bucket, and uploads and downloads bypass the backend entirely through presigned URLs. The backend's job is access policy, routing, identity, and groups. The heavy lifting, encryption, decryption, verification, stays on the client.
DeepSecret isn't trying to replace Vault or a password manager. Those solve different problems, and they solve them well. This is the layer between them: the operational data-exchange channel that teams, and increasingly, machines and AI agents, actually use day to day to move secrets and files around.
That last part is the piece I think gets worse, not better, from here. LLM workflows are another surface where credentials get pasted. Agents running on arbitrary servers need ephemeral, scoped access to do useful work. The number of non-human identities that need to securely receive a secret is going up fast, and Slack threads aren't a plan.
If you've ever pasted some credentials into a DM and immediately wished you hadn't, that's who I built this for.