mcp-modelmanager MCP
A community-hosted MCP server that allows AI agents to manage local model machines running Ollama or vLLM. It supports reading system state (GPU memory, disk, services), pulling and removing models, switching inference services, and performing LoRA training. It is designed for self-hosted, privacy-first local AI environments.
Package and access
Local package with source-attributed installation metadata. Koskamo does not execute packages during review.
What you can do with it
- Read GPU memory and disk state of local model machine
- Pull and remove models from local registry
- Switch vLLM inference service to a different model
- Run LoRA training on custom Ollama variants
Source and maintenance
- Maintenance checked
Packages and installation
Source-provided package data. Koskamo has not executed or live-probed this package.
mcp-modelmanager 1.0.6
{
"version": "1.0.6",
"transport": "stdio",
"identifier": "mcp-modelmanager",
"registryType": "pypi",
"environmentVariables": [
{
"name": "MM_VM_HOST",
"format": "string",
"isRequired": true,
"description": "Address of the model machine (127.0.0.1, LAN IP or tunnel/VPN address)"
},
{
"name": "MM_VM_USER",
"format": "string",
"isRequired": true,
"description": "SSH user on the model machine"
},
{
"name": "MM_CONTAINER_ROOT",
"format": "filepath",
"isRequired": true,
"description": "Root directory on the machine holding all model, adapter and cache data"
},
{
"name": "MM_MAX_TIER",
"format": "string",
"default": "read",
"description": "Highest tool tier exposed: read, control or destructive"
},
{
"name": "MM_ENABLED_OPERATIONS",
"format": "string",
"description": "Optional comma-separated allowlist of exact tool names, applied on top of MM_MAX_TIER; a typo aborts startup. Empty means everything within MM_MAX_TIER"
},
{
"name": "MM_VM2_HOST",
"format": "string",
"description": "Optional: address of a second machine sharing the GPU, used to interpret unreachability. Leave empty if there is only one machine"
}
]
}Tools
Reported by the attributed source; not independently observed by Koskamo.
Connect from your client
Each snippet configures the same source-provided package; replace placeholders before use.
Cursor
{
"mcpServers": {
"mcp-modelmanager-mcp": {
"command": "uvx",
"args": [
"mcp-modelmanager==1.0.6"
],
"env": {
"MM_VM_HOST": "<YOUR_MM_VM_HOST>",
"MM_VM_USER": "<YOUR_MM_VM_USER>",
"MM_CONTAINER_ROOT": "<YOUR_MM_CONTAINER_ROOT>",
"MM_MAX_TIER": "<YOUR_MM_MAX_TIER>",
"MM_ENABLED_OPERATIONS": "<YOUR_MM_ENABLED_OPERATIONS>",
"MM_VM2_HOST": "<YOUR_MM_VM2_HOST>"
}
}
}
}Cline
{
"mcpServers": {
"mcp-modelmanager-mcp": {
"command": "uvx",
"args": [
"mcp-modelmanager==1.0.6"
],
"env": {
"MM_VM_HOST": "<YOUR_MM_VM_HOST>",
"MM_VM_USER": "<YOUR_MM_VM_USER>",
"MM_CONTAINER_ROOT": "<YOUR_MM_CONTAINER_ROOT>",
"MM_MAX_TIER": "<YOUR_MM_MAX_TIER>",
"MM_ENABLED_OPERATIONS": "<YOUR_MM_ENABLED_OPERATIONS>",
"MM_VM2_HOST": "<YOUR_MM_VM2_HOST>"
}
}
}
}VS Code
{
"servers": {
"mcp-modelmanager-mcp": {
"type": "stdio",
"command": "uvx",
"args": [
"mcp-modelmanager==1.0.6"
],
"env": {
"MM_VM_HOST": "<YOUR_MM_VM_HOST>",
"MM_VM_USER": "<YOUR_MM_VM_USER>",
"MM_CONTAINER_ROOT": "<YOUR_MM_CONTAINER_ROOT>",
"MM_MAX_TIER": "<YOUR_MM_MAX_TIER>",
"MM_ENABLED_OPERATIONS": "<YOUR_MM_ENABLED_OPERATIONS>",
"MM_VM2_HOST": "<YOUR_MM_VM2_HOST>"
}
}
}
}Claude Desktop
{
"mcpServers": {
"mcp-modelmanager-mcp": {
"command": "uvx",
"args": [
"mcp-modelmanager==1.0.6"
],
"env": {
"MM_VM_HOST": "<YOUR_MM_VM_HOST>",
"MM_VM_USER": "<YOUR_MM_VM_USER>",
"MM_CONTAINER_ROOT": "<YOUR_MM_CONTAINER_ROOT>",
"MM_MAX_TIER": "<YOUR_MM_MAX_TIER>",
"MM_ENABLED_OPERATIONS": "<YOUR_MM_ENABLED_OPERATIONS>",
"MM_VM2_HOST": "<YOUR_MM_VM2_HOST>"
}
}
}
}Let your agent find its tools
Find servers like mcp-modelmanager MCP from your agent. Add Koskamo once to search the index for endpoints, access requirements, and client setup. Public discovery needs no key.