# setheerwagen/mcp-modelmanager [Health: Active]

**Category:** 🗣️ Conversational AI  
**Repository:** https://github.com/setheerwagen/local-ai-mcp-servers/tree/main/mcp-modelmanager  
**GitHub Stars:** 0  
**Views:** 3  
**Installs:** 0  
**Upvotes:** 0  
**Directory Page:** https://allmcps.com/mcp/setheerwagen-mcp-modelmanager

## Description
Manage a local model machine: pull and remove Ollama models, switch the inference service between models and run LoRA fine-tuning jobs.

## Claude Desktop Quick Installation
Install path detected from listing signals. Uses `uvx` (confidence: high):

```json
"mcpServers": {
  "mcp-modelmanager": {
    "command": "uvx",
    "args": ["mcp-modelmanager"],
    "env": {
      "MM_ACCESS": "",
      "MM_VM_HOST": "",
      "MM_VM_USER": "",
      "MM_CONTAINER_ROOT": ""
    }
  }
}
```

**Requires environment variables:** `MM_ACCESS`, `MM_VM_HOST`, `MM_VM_USER`, `MM_CONTAINER_ROOT` — the values above are empty placeholders; fill in real credentials before running (see the repository for what each one is for).

## Documentation

## What setheerwagen/mcp-modelmanager MCP server does

The setheerwagen/mcp-modelmanager MCP server gives an MCP client controlled access to a local model machine. It can read machine state such as GPU memory, disk usage, the active service, and containers. It also supports managing model files, changing the model used by a vLLM inference service, creating custom Ollama variants, and launching LoRA fine-tuning jobs.

This package is the management half of a two-server repository. The separate `mcp-ollama-vllm` server is intended for sending language requests to a local model; setheerwagen/mcp-modelmanager MCP server is intended for administration. Keeping those functions separate avoids giving a model-querying bridge the same machine-management access.

## How it works

The server speaks Model Context Protocol and is launched as a local command by an MCP client. It includes both the MCP-facing code and machine-side scripts, wrappers, job runners, and network-fence templates. Operations use fixed templates rather than unrestricted remote shell access, and parameters are checked against an allowlist on both sides.

Model retrieval is an explicit operation. `fetch_model` can ask the model machine to obtain a model through Ollama or directly from Hugging Face, depending on the selected source. With the default Ollama source, plain names use the Ollama registry, while prefixed names can identify another registry. `switch_service` does not download models; it refuses a switch when the requested model is not already present and directs the caller to fetch it first.

The server is designed for a single operator and self-hosted hardware. Its own traffic is limited to the local computer or private network, and startup validation rejects public addresses. The MCP client or hosted AI product driving it may still receive the prompts, arguments, and results involved in an MCP call.

## Setup and configuration

Install the package from PyPI with `pip install mcp-modelmanager`. The package provides an `mcp-modelmanager` console command that can be used directly as an MCP client's command. Python 3.11 or newer is required for this server. The repository also provides a source-based setup path and separate documentation for preparing the model machine.

Configuration requires the machine access settings shown in the project documentation: `MM_ACCESS`, `MM_VM_HOST`, `MM_VM_USER`, and `MM_CONTAINER_ROOT`. These select the access mode, target host, SSH user, and container root. Use values appropriate to the local deployment rather than copying example placeholders. No account, paid service, or cloud deployment is required.

## Tools and capabilities

- Inspect GPU memory, disk space, running services, and containers.
- Fetch and remove local models.
- Switch a vLLM inference service between models already available on the machine.
- Build custom Ollama model variants.
- Start LoRA fine-tuning jobs.
- Apply allowlists and fixed execution templates to management operations.

## Limitations and notes

setheerwagen/mcp-modelmanager MCP server is a machine-management interface, not a general-purpose model chat server. It does not replace the companion `mcp-ollama-vllm` package for writing, summarization, classification, extraction, or schema-validated JSON generation.

Fetching a model is the exception to the server's otherwise local traffic pattern: when `fetch_model` is invoked, the model machine connects to the selected public registry. Installing dependencies also requires internet access during setup. Service switching itself does not contact a registry.

The repository describes the server as actively used by its author and version 1.0.0. The project is licensed under MIT. Because the tool can manage models, services, containers, and training jobs, it should be connected only to a machine and operator context where those permissions are appropriate.

_Full upstream README: https://allmcps.com/mcp/setheerwagen-mcp-modelmanager/readme_