Overview

Add a DockerModel provider to plugin-ai that routes inference requests to a locally running Docker Model Runner instance.

Docker Model Runner (docs) exposes an OpenAI-compatible REST API at http://localhost:12434/engines/v1, requiring no authentication. This means DockerModel can extend the existing OpenAICompliantProvider base class and re-use the full provider infrastructure with minimal new code.

Motivation

Docker Model Runner is now a first-class Docker feature available out-of-the-box on Docker Desktop and Docker Engine (Linux). For Kestra users who run local or on-prem workflows, it provides a zero-friction path to self-hosted LLM inference without managing a separate Ollama or LocalAI stack. Adding a dedicated provider makes the integration explicit and discoverable.

Link with plugin-docker

plugin-docker (kestra-io/plugin-docker#119) adds dedicated tasks (docker.model.Pull, docker.model.List, etc.) to manage the Docker Model Runner model lifecycle from a Kestra flow.

The separation of concerns is intentional: plugin-docker owns model lifecycle (pull, configure, delete), plugin-ai owns inference. A typical flow chains both:

tasks:
  - id: pull_model
    type: io.kestra.plugin.docker.model.Pull
    model: ai/smollm2

  - id: ask
    type: io.kestra.plugin.ai.completion.ChatCompletion
    provider:
      type: io.kestra.plugin.ai.provider.DockerModel
      modelName: ai/smollm2
    messages:
      - role: user
        content: What is Docker Model Runner?

Implementation

New class: io.kestra.plugin.ai.provider.DockerModel

Extends OpenAICompliantProvider (same parent as OpenAI, AzureOpenAI).

Properties:

Supported capabilities:

• chatModel() — chat completions (/engines/v1/chat/completions)
• embeddingModel() — text embeddings (/engines/v1/embeddings)
• imageModel() — image generation via Diffusers (/engines/diffusers/v1/images/generations, requires a diffuser-tagged model such as ai/stable-diffusion)

Behaviour notes:

• No API key is required by Docker Model Runner; the apiKey field should default to "not-needed" and be marked optional in the schema
• Base URL for containers should be http://model-runner.docker.internal/engines/v1 (Docker Desktop) or http://172.17.0.1:12434/engines/v1 (Docker Engine)
• The baseUrl field should be prominently documented with both variants

Acceptance Criteria

• DockerModel provider implemented extending OpenAICompliantProvider
• baseUrl defaults to http://localhost:12434/engines/v1; clearly documented with container-internal variants
• apiKey is optional, defaults to "not-needed"
• chatModel(), embeddingModel(), and imageModel() implemented (image generation routes to /engines/diffusers/v1/)
• @Schema, @Plugin, and @Example annotations with at least one working flow example
• Unit tests covering property rendering, default URL, and optional API key
• Integration tests annotated @DockerModelRunnerTest (skippable when DMR is unavailable)

References

• Docker Model Runner docs
• OpenAI-compatible base URL: http://localhost:12434/engines/v1
• Anthropic-compatible base URL: http://localhost:12434/anthropic
• Ollama-compatible base URL: http://localhost:12434
• Supported model list: ai/* namespace on Docker Hub (74+ repos)
• Existing similar provider for reference: io.kestra.plugin.ai.provider.Ollama, io.kestra.plugin.ai.provider.LocalAI
• Companion issue: kestra-io/plugin-docker#119 — model lifecycle management tasks