A secrets manager
built for AI workloads.
An AI secrets manager stores credentials for apps and agents, and delivers them in ways a model can't leak. seekrit is built as one: end-to-end encrypted storage, value-free agent interfaces, and brokering at the network boundary — the model handles names and placeholders, never values.
What makes a secrets manager an AI secrets manager?
The storage half is the familiar job. What changes is the consumer: a stochastic system that will sometimes repeat whatever it can read. Four requirements follow.
The vendor can't read your secrets
Why it's different with AI: An agent-heavy stack multiplies who and what touches the vault. Server-side decryption means every one of those paths trusts the vendor.
How seekrit does it: End-to-end encryption in the browser and CLI; the service stores ciphertext. A full database dump reveals no secret values.
Encryption model →Agent interfaces are value-free
Why it's different with AI: A model repeats what it reads — into transcripts, logs, and tool calls. An MCP tool that returns plaintext is a leak with a schema.
How seekrit does it: The hosted MCP server and agent plugin expose management by name only; no tool returns a secret value. Decryption happens locally, at the moment of use.
MCP servers →Credentials are brokered, not held
Why it's different with AI: Prompt injection turns any credential the agent holds into an exfiltration payload. The fix is structural, not a better system prompt.
How seekrit does it: Agents send {{seekrit:NAME}} placeholders; the proxy substitutes real values at the network boundary against a default-deny allowlist.
Credential brokering →Machines are first-class users
Why it's different with AI: Agents, CI, and infra-as-code outnumber humans in an AI stack, and none of them can complete an email round-trip.
How seekrit does it: Self-serve agent signup over MCP, machine credentials, SDKs in four languages, Terraform, Kubernetes via ESO — and a synchronous audit trail behind all of it.
Agent access →How is this different from Vault or Doppler?
Traditional secrets managers were built for predictable consumers — servers, CI, deploy pipelines — and most of them decrypt server-side because their features need to. That trade is defensible when the consumer is a pipeline. It gets harder to defend when the consumer is an agent, the credential count multiplies, and the vendor's read access sits behind every one of them. seekrit keeps decryption on your side and keeps the agent's side of the interface value-free.
The specifics, claim by claim: seekrit vs Doppler and seekrit vs Infisical agent-vault.
Can an AI agent set itself up?
Yes — an agent can self-register against the hosted MCP server and start organizing secrets by name, with no email round-trip to die on. The hosted server is metadata-only by design: no tool it exposes can return a secret value, so a self-provisioned agent can manage secrets it can never read. Values stay on the crypto side — your machines, your CLI — where a human granted them.
Try it before the next key leaks
Free for one developer, no card. End-to-end encrypted from the first secret.