Vercel Connect is generally available, with the ecosystem past 100 connectors after public beta, unified behaviour and added governance features. The mechanism is narrow and specific: instead of provisioning a long-lived token, code requests a token at runtime, scoped to the task, that expires on its own.
The argument against the usual alternative is the sharper part. Managing credentials became its own workload — rotation scripts, secrets copied across environments, tokens shared between users — and putting a token in a vault made it harder to steal without making it less dangerous once stolen, because it still never expires and no vault limits what a leaked credential can do. Agents made this worse by reaching beyond the codebase to post to Slack, open pull requests, query a warehouse or call an internal API, which is simultaneously the source of their value and of the exposure.
Short-lived and scoped is the same conclusion reached from the identity side in ChainDrop showing agent security starts with identity and in KnownAgents treating bot identity as an authentication boundary. The agent security hub tracks credential scope as a first-class control.
This is a general-availability announcement describing the model and connector count, without published token lifetimes, scope granularity or audit format. The condition to watch is whether Connect exposes which task a token was issued for in an inspectable log, because task-scoped is only auditable if the task is recorded alongside the grant.