Star
Build

Change control

Every way to make a flag change safer: preview it, get it approved, schedule it, undo it, and prove what happened. All of it is in the MIT build, not an enterprise tier.

One change, start to finish

  1. Preview it on real users Dry run: who would change, before anyone does.

    curl -X POST $FLAG/dry-run -H "Authorization: Bearer $FEATUREKIT_TOKEN" -H "Content-Type: application/json" \
            -d '{ "patch": { "enabled": true }, "contexts": [{ "key": "u-ana" }, { "key": "u-ben", "plan": "pro" }] }'

    Send up to 10,000 real contexts. Nothing is saved. Readers can run it.

  2. Ask for a review Change request: a second person approves.

    curl -X POST $FLAG/requests -H "Authorization: Bearer $FEATUREKIT_TOKEN" -H "Content-Type: application/json" \
            -d '{ "patch": { "enabled": true }, "comment": "QA signed off" }'

    Anyone with reader can raise one, which is how someone without write access proposes a change. A writer approves or rejects it. Nobody can approve their own.

  3. Pick the minute Scheduled change: applied by the scheduler, audited like any write.

    curl -X POST $FLAG/schedule -H "Authorization: Bearer $FEATUREKIT_TOKEN" -H "Content-Type: application/json" \
            -d '{ "patch": { "enabled": true }, "executeAt": "2026-11-27T09:00:00Z" }'

    Validated when you create it, so a typo fails now, not on Saturday. A change that couldn't run within an hour of its time (say the fleet was down) is marked failed instead of firing late.

  4. Ship it Every write carries the version it read.

    Two people editing the same flag can't overwrite each other: the second write gets 409 VERSION_CONFLICT with the current configuration, so they can see what changed.

  5. Prove it happened The audit log is a hash chain.

    curl https://flags.acme.com/v1/admin/projects/acme/audit?flag=new-checkout -H "Authorization: Bearer $FEATUREKIT_TOKEN"
          curl https://flags.acme.com/v1/admin/projects/acme/audit/verify -H "Authorization: Bearer $FEATUREKIT_TOKEN"
          # { "ok": true, "checked": 1284 }

    Each row holds the full before and after, who made it, and how (dashboard, token, scheduler, flags file or MCP). It's written in the same transaction as the change, and chained to the row before it, so editing history makes verification fail at that row.

  6. Undo to any point Revert: back to exactly how it was before an audited change.

    curl -X POST $FLAG/revert -H "Authorization: Bearer $FEATUREKIT_TOKEN" -H "Content-Type: application/json" \
            -d '{ "auditId": "8e2a3b4c-5d6e-4f7a-8b9c-0d1e2f3a4b5c", "comment": "back to 20%" }'

    A revert goes through the same validation as any write, and is audited as flag.reverted.

In these examples $FLAG is https://flags.acme.com/v1/admin/projects/acme/environments/production/flags/new-checkout.

The kill switch

During an incident, skip all of the above:

curl -X POST $FLAG/off -H "Authorization: Bearer $FEATUREKIT_TOKEN" -H "Content-Type: application/json" \
        -d '{ "comment": "INC-2231 checkout erroring" }'
  • No version needed, so you don't have to read the flag first from a control plane that might be struggling.
  • Pushed immediately, skipping the 200ms batching other changes get.
  • Flags that depend on it turn off too: a flag whose prerequisite is off serves its off variation.
  • Still audited.

Who can do what

ActionLowest role
Read flags, the audit log, schedules and requests. Dry-run. Raise a change request.reader
Change a flag, the kill switch, schedule, approve or reject, revert, manage segmentswriter
Verify the audit chain, manage SDK keys, create projects and environmentsadmin
Manage membersowner

Tokens can also be limited to some environments: give CI a token for staging only, and it can't touch production at all.

Retries are safe

Send an Idempotency-Key header on any write, and a retry within 10 minutes returns the first response instead of applying again. Useful for scripts that retry on timeout.