Skip to content
Ambry

Secrets, kept by reference

Keep what matters out of sight.

Ambry is a self-hosted secrets service built for a world where AI agents do the deploying. Agents create, rotate and request secrets without ever holding a credential that can reveal one. People reveal, deploy runners inject, and every open is written to the audit chain.

  • Encrypted at rest, one key per value
  • Agents never see a value
  • Every open is audited
An illustration of the dashboard, restyled for this page; the names and fingerprints are invented. Values stay masked until a person clicks Reveal, which is audited, and they are masked again after 30 seconds.

One rule, enforced twice

Agents manage secrets by reference. They never hold a credential that can reveal a value. The MCP server has no reveal tool, and the server refuses an agent token on reveal and inject, so a clever client does not get around it.

Ambry exists because agents decrypting a vault in a shell printed values into transcripts, and a redaction pipe is not a control. The only fix that holds is an agent credential that cannot return plaintext at all.

What people, agents and deploy runners can each do in Ambry
Action People The dashboard and ambry, logging in each time Agents An MCP server and an agent token Runners A deploy runner and a token scoped to one namespace
List secrets, read their metadata, query the audit log Yes Yes Yes
Create a secret whose value the server generates Yes Yes gets back a fingerprint, never the value No
Rotate a secret to a new generated version Yes Yes same: a fingerprint comes back No
Enter a value only a person has Yes No files a request; a person fulfils it No
Reveal a value Yes an explicit, audited action No refused by the server, whatever the client No
Inject values into a deploy step No No refused by the server Yes only for the namespace it is granted
Administer principals, tokens and grants Yes No No

Used, not handed over

An agent that needs a secret put to work triggers a pipeline. The runner holds the only token that can inject, scoped to one namespace, and the value goes into the step. The agent sees the step's output.

  1. Step 1

    The agent

    Needs a secret used. It triggers the deploy pipeline and names the target.

    Holds: An agent token: list, create, rotate, ask. It cannot reveal.

  2. Step 2

    The runner

    Fetches the target’s namespace from Ambry and injects the values into the step.

    Holds: A runner token granted inject on exactly one namespace.

  3. Step 3

    The step

    Runs with the value in its environment. The runner redacts known values from the log.

    Holds: The value, for the length of the step, and nothing that can fetch another.

Built to be boring to attack

libsodium primitives only, one key that never enters the database, and an audit chain that has to accept a row before a value moves.

Agents work by reference

An agent can list, create, rotate and ask, and gets back names and fingerprints. The MCP server has no reveal tool, and the server refuses an agent token on reveal and inject whichever client sends it.

Runners inject, per namespace

A deploy runner holds the only token that can put a value into a step, granted on exactly one project/env. A credential that deploys staging provably cannot touch production.

Versioned, and bound to its place

Every rotation is a new version. Each value is sealed under its own key, and the ciphertext is bound to its namespace, path and version, so a row moved elsewhere fails to open.

Nothing opens unaudited

Every write and every open is recorded in the same transaction as the change, on a chain MACed under a key only the server holds. No audit row, no value.

One key, never in the database

The key-encryption key is a 32-byte file on the host, readable by the service alone. The server refuses to start without it, and refuses a store sealed under a different one.

Backups proved before they are sealed

The nightly job restores its own dump into a throwaway cluster, verifies the audit chain and opens every value, and seals only the bytes that passed, to a key kept offline.

What it does not do yet

Three things are named as not built, among them biometric sign-in, finer reveal permissions, and kek rotation. One more are built and not yet running. The status page lists each one in its own words.

Built, in progress, not built Twelve components are running today

See how the pieces fit

The server, the CLI, the MCP server, the dashboard and the deploy integration, and what each one is allowed to hold.