Custom / self-hosted LLM connections stop working after upgrading to v12 (license enforcement)

Describe the Bug

Summary

Upgrading to Directus 12 removes a previously working capability: an instance configured against a user-supplied, self-hosted LLM endpoint (Ollama / vLLM / LiteLLM / any OpenAI-compatible base URL) loses that connection once license enforcement kicks in.

Per the v12.0.0 release notes, Core-tier self-hosted instances lose:

• Custom or self-hosted LLMs — connections to custom LLMs will no longer work
• AI Translations
• SSO
• Custom permission rules

Enforcement is immediate on new instances; upgraded instances get a 30-day grace period.

Why this one is different from the others

The other gated features have a plausible "this is commercial functionality" story. This one inverts its own purpose: it takes the cheapest, most privacy-preserving, most self-reliant AI configuration and makes it the paid one, while the vendor-routed default presumably stays available.

There is also no vendor cost being recovered here. Blocking a user-supplied base URL and API key is not withholding a service Directus provides — it is an artificial restriction on where a self-hoster may point their own configuration.

Who this breaks

• Air-gapped / on-prem deployments. No egress to a hosted model provider means no AI path at all. Not degraded — none.
• Data-residency and GDPR-constrained setups. Organizations running a local model precisely because they cannot send content to a third-party API are pushed to a commercial license or off Directus AI entirely.
• Cost-controlled deployments. Running your own inference on your own hardware was the main reason many self-hosters enabled these features.
• The "free under $5M revenue / 50 employees" promise. A qualifying team that self-hosts a model is effectively told their own inference endpoint is a paid feature.

Expected behavior

A custom / self-hosted LLM endpoint configured by the operator should keep working on the Core tier, or at minimum under the free OIG license key.

Suggested resolutions

1. Move custom LLM endpoints to Core and keep the gate on Directus-provided AI services (hosted inference, AI translations).
2. Keep the gate, but scope it to the vendor-routed path only — a user-configured base URL + key is not a Directus-delivered feature.
3. At minimum, enable custom LLM connections under the free OIG license so under-$5M teams are not excluded.

Questions

• Was the intent to gate Directus-hosted inference, with custom endpoints caught in the same flag?
• Does the free OIG license key enable custom LLM connections, or is this Enterprise-only?
• Is there a supported way to keep a bring-your-own-model setup working on v12 without a paid license?

To Reproduce

1. On Directus 11.x (self-hosted, no license key), configure a custom / self-hosted LLM provider — e.g. an OpenAI-compatible endpoint such as Ollama or vLLM with a custom base URL.
2. Confirm the AI features using that provider work (e.g. AI-assisted fields / translations / the MCP chat tooling backed by that endpoint).
3. Upgrade the instance to v12.x (e.g. directus/directus:12.3.1) without adding a license key.
4. Wait out the 30-day grace period, or start a fresh v12 instance where enforcement is immediate.
5. Trigger the same AI feature.

Expected: the request is sent to the configured self-hosted endpoint, as it was on v11.

Actual: the connection to the custom LLM is refused by license enforcement. Because the endpoint is self-hosted and possibly on an isolated network, there is no fallback provider available — the feature is simply gone.

Note: filed here because the issue chooser routes feature requests to roadmap.directus.com and questions to community.directus.com. This is a behavioural regression between v11 and v12, so the bug form seemed the closest fit — happy to move it if there's a better venue.

Directus Version

v12.3.1

Hosting Strategy

Self-Hosted (Docker Image)

Database

No response