Quick answer: A secrets broker checks whether a caller may access a credential. A credential-release broker returns the approved key to the authorised runtime and records the retrieval. A request-executing proxy is a different model: it keeps the provider key inside the proxy and applies rules to the requests it forwards. Gazebo currently provides credential retrieval, not provider-request proxying.
Secrets manager vs. secrets broker
“Secrets manager” usually describes the system that stores and distributes secrets. “Broker” describes an access-mediating role in the architecture. They are not mutually exclusive product categories.
Established secrets managers can authenticate machines, apply policies and record access. A separate broker may add a useful interface for agent clients or a shared policy workflow across credential sources, but it should not duplicate controls your existing infrastructure already handles well.
The distinction to evaluate is what happens on a request: who authenticates the caller, which policy is checked, what value is returned and which system controls its later use.
What a credential-release broker does
The retrieval path has four steps:
- Identify the caller. The agent authenticates with its own access token or an authorised connection. That bootstrap credential is distinct from the provider key it wants.
- Check retrieval policy. The broker checks whether the caller may retrieve the requested service credential. Gazebo's declared-method check is policy context, not inspection of a later provider call.
- Return or deny. An authorised request receives the configured stored credential. The runtime and its host receive the actual key; whether it also enters model context or logs depends on the client.
- Record the attempt. The credential-access record describes the requester, requested credential and outcome. It is not a record of subsequent provider API operations.
This can make credential distribution easier to manage without changing the provider key's underlying permissions. Use provider-restricted credentials wherever the issuer supports them.
For the full storage, authentication and runtime design, see the reference architecture.
Retrieval and request proxying solve different problems
A release broker answers “may this agent receive this key?” A proxy can instead answer “may this caller make this particular request?” when it has the necessary endpoint-specific policy and all relevant traffic passes through it.
Only the latter model can keep the provider credential out of the agent runtime while forwarding provider calls on its behalf. It adds operational responsibilities too: provider compatibility, safe request handling, forwarding availability and rules for the actual endpoints being used.
Do not infer proxy capabilities from a broker's support for a method parameter. A GET declaration during credential retrieval cannot make a write-capable key read-only once delivered.
Credential approval is not operation approval
In Gazebo's OAuth connection flow, the user explicitly approves the stored service keys the client may retrieve. This is consent to credential access, not approval of every later action performed with those keys.
For example, approving retrieval of a Stripe credential does not mean Gazebo reviews each later charge or refund. Stripe's credential permissions and any controls in the runtime govern direct requests. Gazebo does not currently provide provider-request inspection or operation-level approval gates.
Revoking retrieval vs. revoking a provider key
With Gazebo, each agent authenticates using its own token or authorised OAuth connection. Revoking that access stops future retrieval without changing the underlying provider credential. Other authorised profiles can remain active.
Previously delivered copies are different. They remain usable within the provider's permissions until expiry or provider-side revocation. If a runtime or key may be compromised, restrict, revoke or rotate that key at its provider as well. Replacing a shared key can affect other consumers, even when their retrieval identities are separate.
A proxy can check authorisation on future forwarded requests, but cannot undo completed operations. It also cannot control use of a provider key obtained through a different route.
For incident and offboarding responsibilities, use the credential-management lifecycle.
When an additional broker is worth considering
Evaluate the actual workflow rather than assuming every agent needs another service:
- One trusted local agent: Existing runtime handling and a narrowly scoped provider key may be enough.
- Several independently managed agents and services: A central connection, retrieval-approval and access-history workflow may reduce manual distribution and make reviews easier.
- An established platform team: Existing workload identities and secrets policies may already meet the need. An added broker should prove that integration becomes easier.
- A requirement to keep provider keys away from runtimes: Stored-key retrieval does not satisfy that requirement; evaluate a request proxy or another architecture.
For a managed service, assess runtime trust, availability, key handling, audit coverage, exportability and pricing. For an internal implementation, assess who owns authentication, policy changes, secret redaction, recovery and ongoing support. Neither choice is automatically safer or cheaper.
Existing vaults, performance and evidence
A broker can sit in front of an existing vault if the implementation supports it. For example, an internal service might authenticate to Vault through AppRole or use an IAM role to retrieve an AWS Secrets Manager value, then apply its own retrieval policy. These are architectural examples, not a claim that every broker supports every source or that a returned key becomes narrower.
Check the supported credential-source options before choosing a Gazebo integration. Retain the underlying vault's authentication and access controls rather than treating the broker as a replacement for them.
Measure retrieval latency and availability for the workflow you intend to run. Caching reduces repeat retrieval but allows a runtime to retain a valid key after broker access is revoked. A retrieval outage therefore need not stop ongoing provider calls.
Retrieval records can support access reviews and investigations, but they do not establish provider activity or compliance by themselves. Keep separate evidence for policy approvals, review decisions, provider operations and provider-key invalidation where required.
What this looks like with Gazebo
Gazebo is a credential-release broker built for AI agents. Store service credentials, then give each agent a named identity and a policy controlling which credentials it may retrieve and under which declared methods. Create and restrict provider credentials at the provider where supported; Gazebo does not automatically mint narrower provider keys.
When an agent needs a credential, it calls Gazebo through MCP or the API using its agent token or an authorised OAuth connection. Gazebo checks the credential-retrieval policy, returns the approved stored provider credential to the runtime, and records the retrieval attempt. The runtime and its host receive the actual key; whether it enters model context depends on the client.
Gazebo does not proxy subsequent provider requests or enforce how a retrieved key is used. Its logs record credential retrieval, not downstream provider actions. Revoking access stops future retrieval; provider-side permissions, expiry and revocation govern keys already delivered.
The SDK and CLI provide additional integration paths. Use the six-practice checklist to verify the resulting setup rather than assuming the broker alone supplies every control.