Separate lifetimes
A request has a short time to be answered: 15 minutes by default, set by your server between 1 minute and 24 hours. If nobody answers, it expires and nothing is granted.
How long access lasts is separate. The approver chooses it, and it can be shorter or longer than the requester asked for, within the resource's limit. The longest custom length a server allows is 365 days.
From the source; the test above revoked the grant rather than waiting for it to expire. When a grant ends, Bouncer stops issuing tokens for it and tries to revoke those already issued; failed revokes are retried. An issued token can remain usable until revocation succeeds or it expires. A token's lifetime is capped by the grant's remaining time. Expiring a token does not expire the stored password or API key it read. Change or disable a copied value at its source. Bouncer does not end an established session; an SSH session, for example, can outlast its certificate.
Standing access
Servers can offer Always, a standing grant that stays until it is revoked or its review date passes (every 90 days by default). It is a revocable approval, not a credential that never expires. It is allowed in the default config, but its policy isn't settled yet, so we recommend turning it off (durations.always: false) and removing it from the quick choices.
Where it runs
Everything except notification delivery runs on infrastructure your organisation controls. A typical setup puts the server on a private network, such as Tailscale, behind HTTPS. Requesters need network access to OpenBao in every delivery mode: the CLI and MCP read values or unwrap responses there themselves.
What the records show
Bouncer's approval log records requests, decisions, grant revocations, app views and saves, sharing changes, device pairing and revocation, the kill switch, and device-reported screenshots and copies. The server separately tracks issued tokens for revocation. OpenBao's audit device records requests made to OpenBao when that audit device is enabled. The database, SSH host or other target needs its own logs for what someone did after connecting. Bouncer does not proxy or record those sessions.
Administrators can export Bouncer's approval log as JSON Lines with bouncer-admin log export. Bouncer doesn't copy the log off the server; set that up yourself. Keep the Bouncer, OpenBao and target-system records together for an investigation. A device report about a screenshot or copy is not independently verified, and a missing report doesn't mean nothing was captured.
Matching the records in a test run
In a separate run on the same build, also on 17 November 2026, we followed one fake request from build-agent through approval (by Alice, on the same test helper), a successful read and revocation, then matched up the records.
The exported approval log links request req_01m7n3gfz272rt3yzp8cja1bjx to grant grt_01m7n3gg0msygady2dm8jd3ync. OpenBao's audit entry for the read carries the same grant ID in auth.metadata.bouncer_grant, and an issuance ID in bouncer_issuance that matches the one in Bouncer's server log and revoke response.
| Step | Observed record, UTC | What matches |
| Request and approval | Exported approval log: request filed and approved at 15:55:17. | The same request and grant appear in the approval and later revocation records. |
| Issue and read | Bouncer server log: token issued at 15:55:21. OpenBao audit: a successful read of kv/data/apps/billing/db-password immediately afterwards. | The read's grant ID matches the approval export. Its issuance ID matches the server log. The token arrived wrapped, as it does by default; the requester exchanged the wrapping token at OpenBao before reading. |
| Revoke | Exported approval log: grant revoked at 15:55:21. OpenBao confirmed token revocation that second. | Bouncer's revoke and OpenBao's token revocation are separate records. A later bouncer exec --no-request exited 8 without running its child. |
An extra check, in the test setup only. We also checked that the revoked token itself stopped working. You can't repeat this with Bouncer's tools. Bouncer's exported approval log has no token accessor (OpenBao's reference ID for a token), and the server log shows only the first few characters, which isn't unique. So the test captured the token and its accessor from OpenBao's unwrap response. The token stayed on the disposable test machine and was removed afterwards. OpenBao's audit log stores keyed hashes (HMACs) of tokens and accessors rather than the values, so we used OpenBao's audit-hash API on the same audit device to find them in the log. The accessor matched the entries for creating the wrapped token, the read and the revoke. A read with the same token after the revoke, before its 120-second lifetime ran out, was refused with HTTP 403. That entry had no accessor, but its token hash matched the successful read. Response wrapping, policy and the 120-second token lifetime stayed on.
These records don't show what the child command did with the value, any later use of a copied value, or activity in a session. No database, target connection or target-side log was part of the test; those systems need their own logs. Bouncer doesn't join the records for you; you compare them yourself.
Where Bouncer fits beside HashiCorp Vault and Boundary
Bouncer sits in front of OpenBao. It adds a request that a person reads, a decision signed on their device, and a grant with an end time. It is built for OpenBao and hasn't been tested with HashiCorp Vault.
Human approval is also available in Vault Enterprise control groups, which require extra approvals before Vault answers a request. Availability depends on the Vault edition and licence; see HashiCorp's edition comparison. We don't claim Bouncer matches control groups feature for feature.
Boundary controls connections through workers. Its session recording supports SSH and RDP targets when configured on the Plus and Enterprise tiers. Bouncer controls credential issuance rather than the ongoing connection, and its approval log is not session recording. Check HashiCorp's current editions and protocol support against your needs.
Read the security model for what this protects against and where it stops.