secr vs Doppler, Infisical, and Vault — for AI agents

If your agents are well-behaved, internal, low-stakes: any of the four work. Inject an API key, run the agent, move on. The differences below only start to matter when something goes wrong — a misaligned model, a prompt injection, a token in a public repo, an agent that calls a tool it shouldn't.

If you're already running Vault for your platform: you can wire Vault policies tight enough to scope an agent to a few paths. You can't naturally do tool-call governance or approval queues — that's outside Vault's model. Most teams run secr alongside Vault: Vault for the platform, secr for the agent layer.

If Doppler or Infisical is enough: stay there. They are excellent products for what they do. The case for adding secr is specifically when you have AI agents in the loop and want the gate to be in the credential layer, not in the agent code.

What secr does that the others can't emulate easily: the per-agent allowlist is a primitive of the data model, not a policy you write. The MCP gateway runs in-line with the agent — the before_tool_call hook fires before the network call, so blocking is real, not a tripwire. The approval queue uses Postgres FOR UPDATE SKIP LOCKED so concurrent retries can't double-spend a single approval — you'd need to write this protocol yourself in any of the other tools.