Vercel's free Glyph Cluster model may hand your prompts to an unnamed provider for training

AI Gateway lists Zero Data Retention support among its built-in features ^[13]. Vercel's October 7 changelog makes Glyph Cluster the exception: ZDR is not available for it ^[11][2]. The changelog does not say how the gateway feature treats a request routed to a model without ZDR.

The training signal has a second wrinkle. The dev.to plan's author treats the changelog warning as the operative constraint and recommends asking Vercel before relying on a different reading ^[6]. Read literally, a Stealth row without a "No Training" mark agrees with the changelog. Either way, the changelog is the strictest published wording, and testing should follow it ^[2].

The plan also warns that stealth access and its terms can change ^[9]. Approval therefore covers an unnamed provider ^[14] on terms that can move after sign-off.

The dev.to plan draws the test boundary tightly. Only public material, or non-sensitive examples the team owns and may send, goes to the model. Private repositories, customer data, access tokens, incident reports, unreleased designs and real support conversations stay out ^[7]. Redaction does not qualify, because snippets can keep context that identifies a customer or exposes implementation details ^[7]. "A model that cannot be safely evaluated for your workload is not a candidate for that workload yet," the author wrote ^[8]. I'd keep that cut, even though private code heads the excluded list for a model Vercel sells on coding ^[7][3].

The quickest integration runs straight through it. Vercel's coding-agent setup is a CLI install, `vercel ai-gateway setup`, and a selection of `stealth/glyph-cluster` inside Claude Code, Codex or Cursor ^[5]. An agent pointed at a private repository is exactly the case the plan excludes ^[7]. Its alternative is a development-only route or a local script. The key stays server-side in an environment variable, never in a browser bundle or a committed file ^[16].

Function calling does not close the structured-output gap, the plan notes ^[12]. It keeps payments, permission decisions, automatic customer-facing actions and schema-validated flows out of the first round ^[15]. Its starter tasks are explaining a public build failure, comparing two owned implementation plans, and patching a small sample repository with no secrets ^[15]. Expected outcomes are written before either model runs. They include a seeded bug, a clean sample and a case where the right answer is to ask for more context ^[17].

A pass on that set is a result about a sample repository. It transfers to private code only as far as the sample resembles the code a team ships, and the boundary rules out checking that against the real thing. The author wrote that free pricing should not be what gates a move to production ^[10].