[MuhammadUsamaMX]

Problem

Per-model blocking already exists, but it is only reachable from one provider's page at a time, and there is no single place to see what the final endpoint actually serves. With a number of providers connected, /v1/models ends up listing models that were never deliberately curated, and auditing that list means walking every provider page in turn.

What this adds

A Model Access page (/dashboard/model-access, sidebar entry under Providers) that lists every model of every connected provider, split into allowed and blocked:

• per-model allow/block toggle
• per-provider Block shown / Allow shown, operating on the currently filtered set — so you can search a term and act on just those
• search across model id and provider, plus All / Allowed / Blocked filters
• allowed / blocked / total counts per provider and overall

A Fetch Models button on the provider page, for the providers that did not already have one. OpenAI-compatible providers keep their existing Import from /models; this gives the others the same ability to list what the provider itself reports, using an active account's credentials (/api/providers/<id>/models) rather than only the static registry. Allowed/blocked count badges were added to the same header.

Approach

This deliberately does not introduce a second gating mechanism. /api/models/access reads and writes the same per-provider disabled list (disableModels / enableModels) that /v1/models already filters on, so a model blocked here genuinely disappears from the endpoint and stays blocked across restarts.

New endpoint:

• GET /api/models/access → { providers: [{ alias, connections, total, allowed, blocked, models[] }], totals }
• POST /api/models/access → { providerAlias, ids[], action: "block" | "allow" }

Only providers with a connected account are listed, since a model from a provider with no account can never be served anyway.

Testing

Built and run against a live instance with several providers connected (86 models across 5 providers). Round-trip verified against the real endpoint:

• block a model → it disappears from GET /v1/models
• allow it again → it comes back
• blocked counts return to their prior state, no residue left behind

docker build is clean on top of master.

Related

• Feature request: Per-model disable/enable toggle #1525 — asks for a per-model disable/enable toggle. The per-model disable itself already exists; this adds the central place to see and drive it across every provider at once.
• Custom third-party provider: /v1/models exposes ALL upstream models instead of only Available Models #2768 — /v1/models exposing all upstream models rather than only curated ones. This does not change that default behaviour, but it does give a practical way to review and block the models that get surfaced.