One Endpoint, Three SDKs

A minimal demo showing how Docker Model Runner (DMR) exposes the same local model through three native API formats (OpenAI, Anthropic, Ollama) on a single endpoint, and how that unlocks dev/prod parity with Microsoft Foundry.

🇪🇸 Versión en español

Why this matters

Most local LLM runtimes speak one protocol. DMR speaks three, natively, on the same port. That means:

• One model in memory, consumed by the OpenAI SDK, the Anthropic SDK, or any Ollama-compatible client.
• No translation proxy, no LiteLLM in the middle, zero extra hops.
• Same SDK code in development (against local DMR) and production (against Microsoft Foundry or Anthropic API). One env var changes.

This repo proves it with six short, runnable examples.

Prerequisites

• Docker Desktop 4.58.0+ (Anthropic API support requires this version)
• Python 3.10+ (for the Python examples)
• .NET 10 SDK (for the C# examples — dotnet run from the examples/dotnet project)
• GPU optional, llama.cpp runs on CPU and Apple Silicon

Setup

1. Enable Docker Model Runner

# Enable DMR with TCP access on port 12434
docker desktop enable model-runner --tcp 12434

# Verify it is running
curl http://localhost:12434/engines/v1/models

2. Pull a model

Pick one. The examples default to ai/qwen3-coder (good for coding, 128K context).

docker model pull ai/qwen3-coder
# Alternatives:
# docker model pull ai/llama3.2          # smaller, faster
# docker model pull ai/devstral-small-2  # coding-focused
# docker model pull ai/mistral           # general purpose

3. Install Python dependencies

python -m venv .venv
source .venv/bin/activate          # Windows: .venv\Scripts\activate
pip install -r requirements.txt

4. Configure environment

cp .env.example .env
# Edit .env if you want to test the production-parity example

Run the examples

01_openai_local.py — Local model via OpenAI SDK

python examples/01_openai_local.py

DMR exposes /engines/v1/chat/completions. The OpenAI Python SDK sees it as a regular OpenAI endpoint. No translation.

02_anthropic_local.py — Same model via Anthropic SDK

python examples/02_anthropic_local.py

DMR exposes /v1/messages. The Anthropic Python SDK sees it as a regular Anthropic endpoint. Same model, same memory, just a different protocol.

03_streaming.py — Streaming with both SDKs

python examples/03_streaming.py

Token-by-token streaming works identically on both protocols.

04_dev_prod_parity.py — The killer pattern

# Development: hit local DMR
ENVIRONMENT=development python examples/04_dev_prod_parity.py

# Production: hit Microsoft Foundry (requires .env values)
ENVIRONMENT=production python examples/04_dev_prod_parity.py

Same client code, same SDK calls. Only the base_url and credentials change.

05_curl_examples.sh — Raw API calls

bash examples/05_curl_examples.sh

Hit the three protocols directly with curl. Useful for debugging and for understanding what each SDK sends under the hood.

06_ollama_local.py — Same model via Ollama SDK

python examples/06_ollama_local.py

DMR exposes /api/chat. The Ollama Python SDK sees it as a regular Ollama server. Same model, same memory, third protocol. This completes the picture: one model in memory, three Python SDKs, zero proxies.

.NET 10 / C# examples

The same five examples are available as a single .NET 10 console project under examples/dotnet/.

Packages used:

SDK	NuGet package	Version
OpenAI	OpenAI	2.10.0
Azure OpenAI (Foundry)	Azure.AI.OpenAI	2.1.0
Anthropic (official)	Anthropic	12.17.0
Ollama	OllamaSharp	5.4.25

Run:

# Build once
dotnet restore examples/dotnet

# Then pick any example
dotnet run --project examples/dotnet -- 01   # OpenAI SDK
dotnet run --project examples/dotnet -- 02   # Anthropic SDK (official)
dotnet run --project examples/dotnet -- 03   # Streaming (both SDKs)
dotnet run --project examples/dotnet -- 04   # Dev/prod parity
ENVIRONMENT=production dotnet run --project examples/dotnet -- 04
dotnet run --project examples/dotnet -- 06   # Ollama SDK

The project reads the same .env file at the repo root via DotNetEnv.Env.TraversePath().Load(), so no extra configuration is needed.

What the demo proves

Pattern	Demonstrated by
One model, three SDKs, no proxy	01, 02, 06
Streaming parity across protocols	03
Dev/prod parity with Foundry	04
Raw protocol inspection	05

Architecture

┌─────────────────────────────────────────────────────────┐
│                    Your laptop                          │
│                                                         │
│  ┌──────────────┐      ┌────────────────────────────┐   │
│  │  OpenAI SDK  │─────▶│ /engines/v1/chat/completions│   │
│  └──────────────┘      │                            │   │
│                        │   Docker Model Runner      │   │
│  ┌──────────────┐      │                            │   │
│  │ Anthropic SDK│─────▶│       /v1/messages         │   │
│  └──────────────┘      │                            │   │
│                        │   ┌──────────────────────┐ │   │
│  ┌──────────────┐      │   │ ai/qwen3-coder       │ │   │
│  │ Claude Code  │─────▶│   │ (one model, in memory)│ │   │
│  └──────────────┘      │   └──────────────────────┘ │   │
│                        └────────────────────────────┘   │
│                              localhost:12434            │
└─────────────────────────────────────────────────────────┘
                              │
                              │ same SDK calls, swap base_url
                              ▼
                   ┌──────────────────────┐
                   │   Microsoft Foundry  │
                   │   (production)       │
                   └──────────────────────┘

Endpoint reference

Protocol	Path	Used by
OpenAI Chat Completions	/engines/v1/chat/completions	openai, langchain-openai, AI SDK
OpenAI Responses	/engines/v1/responses	openai (reasoning models)
Anthropic Messages	/v1/messages	anthropic, Claude Code
Anthropic Token Count	/v1/messages/count_tokens	anthropic
Ollama Chat	/api/chat	Ollama clients

Resources

• Docker Model Runner docs
• DMR REST API reference
• Microsoft Foundry
• Anthropic Messages API

License

MIT

About the author

Pablo Piovano — Director of AI at OZ Digital · Microsoft MVP in AI

---

Built as a companion to a LinkedIn article on #MicrosoftFoundry and local AI development.