EN|中文

A practical first step into local AI.

MTCode Server brings model selection, download, local inference, and secure publishing into one desktop workflow. Its built-in MTCode-LLM and MTCode-Diffusion servers are designed for people who want to get a useful local model running before learning a larger AI tool ecosystem.

Download MTCode Server See the quick start

A guided local AI setup

Choose a task first, then a model that fits the machine

Instead of beginning with model formats, inference flags, and a long catalog, start with an application category. MTCode Server presents a preselected set of open-weight models hosted on Hugging Face, along with quantization and runtime choices intended to make the tradeoff between capability, speed, and GPU memory easier to understand.

Task-based categoriesStart with coding, chat, image understanding, image generation, editing, or inpainting.
Curated model choicesBrowse a smaller set of preselected Hugging Face models instead of an unfiltered catalog.
Hardware-aware optionsChoose model size, quantization, context, and low-memory settings for the available GPU.
Ready to shareStart the local server and publish it through DirectLink from the same application.

Three pieces, one workflow

The GUI manages the model servers and their access

MTCode Server

The desktop controller stores credentials, downloads selected models, configures the local server, starts and stops it, and publishes its TCP endpoint to authorized DirectLink users.

MTCode-LLM

A local language and multimodal model server using the llama.cpp inference engine. It supports the text, coding, vision, and audio-oriented model choices exposed by MTCode Server.

MTCode-Diffusion

A local image generation and editing server using the stable-diffusion.cpp engine. MTCode Server organizes supported workflows into generation, generation-and-editing, and inpainting categories.

Quick start for new users

From an empty machine to a shared AI endpoint.

The built-in path keeps the first setup inside one application. Advanced controls remain available, but they do not have to be the first thing a new user learns.

1. Pick the kind of applicationSelect a language task such as coding or general chat, or an image task such as generation or editing.
2. Choose a model and quantizationReview the curated choices and select a size that fits the local GPU and the quality you need.
3. Download and start locallyMTCode Server retrieves the model from Hugging Face, configures the built-in server, and shows its status and logs.
4. Publish for authorized usersDirectLink makes the endpoint available in MTCode Portal without requiring a public inbound port or VPN.

MTCode-LLM: select an application category, model, quantization, and context before starting the server.Click the image to enlarge.

MTCode-Diffusion: select a generation or editing workflow and a compatible image model.Click the image to enlarge.

Where it fits

Choose it for convenience, not ecosystem breadth

Ollama, LM Studio, and established image-generation tools have broader ecosystems, integrations, catalogs, and community workflows. MTCode-LLM and MTCode-Diffusion are not presented as replacements for those tools. Their advantage is a smaller, guided path from selecting a useful open-weight model to running and sharing it from the MTCode Server GUI.

New to local AI

Use the built-in servers when application categories and a curated shortlist are more helpful than evaluating every runtime and model independently. They provide a practical starting configuration that can be refined later.

Already have a preferred AI server

Keep it. MTCode Server can publish any TCP service, including an existing Ollama, LM Studio, or other compatible endpoint. DirectLink does not require the service to use an MTCode inference engine.

Models remain third-party software and content. Availability, hardware requirements, capabilities, and license terms vary by model. Review the Hugging Face model card and license before downloading or sharing access.

Local by design

Run the model near the data and the GPU.

Inference runs on the selected computer. DirectLink supplies authorized remote access to the server endpoint, while MTCode Portal gives users a stable local address and a familiar client workflow.

Run either server independentlyMTCode-LLM and MTCode-Diffusion are standalone programs. You can launch either one directly from a terminal for local use without signing in to an MTCode account; an account is only needed if you choose to publish the endpoint through DirectLink.
No public listening portPublish the service without configuring router port forwarding or placing the model server directly on the public Internet.
Use the client you preferPortal maps the remote endpoint to localhost, so browser interfaces, API tools, coding assistants, and other compatible clients can connect normally.
Share the service, not the computerUsers receive access to the named AI endpoint rather than general network or operating-system access.

Start with a curated model. Learn the rest when you need it.

MTCode Server provides a direct route from local GPU hardware to a running language or image model—and an equally direct route for authorized users to reach it.

Download MTCode Server Server & Portal Get Started