MCmidnight-coderlocal coding agent

Model and provider boundaries

Use local code LLMs with Midnight Coder

Midnight Coder selects models exposed by a provider; the provider decides which model formats and inference backends it can run.

To use a local coding LLM, expose it through a model server and connect Midnight Coder to that endpoint. The built-in flow checks /api/tags and /v1/models, then lets you select a detected model.

What Midnight Coder supports directly

Midnight Coder supports provider discovery, model selection, reasoning-effort selection, and Ollama smart context. Its documented setup accepts an ip:port or HTTP address and stores the provider and model found at the endpoint.

/provider_conf 127.0.0.1:11434
/model
/smartcontext true

Midnight Coder does not load GGUF files itself. It sends requests to the configured model provider.

GGUF, llama.cpp, Qwen Coder, and DeepSeek Coder

GGUF
A model file format used by some local inference runtimes. Compatibility depends on the provider that loads the file.
llama.cpp
An inference runtime that can serve models. Midnight Coder compatibility depends on exposing a model-list and generation API the configured provider flow can use.
Qwen Coder
A family of coding-focused models. Use is possible only when your chosen provider supports the exact model and exposes it through the connected endpoint.
DeepSeek Coder
A coding-model family subject to the same provider rule: the inference server, not Midnight Coder, determines model-file and hardware compatibility.

Select a code LLM without guessing

To run Qwen Coder locally, or use local coding AI with Ollama and another model family, first make the model work in the provider itself.

  1. Install and start the provider you intend to operate.
  2. Load a model supported by that provider.
  3. Confirm the provider lists the model through its API.
  4. Run /provider_conf and enter the provider address.
  5. Use /model to select only from models Midnight Coder detected.
  6. Start a new chat so the saved provider applies to the full session.

This sequence avoids implying that a model name or file format is supported before the provider proves it.

Connect your model provider

Use Midnight Coder's provider flow to detect the local models your server actually exposes.

Explore Midnight CoderOllama setup