Skip to main content
Gcore MCP Server provides AI and machine learning infrastructure tools that AI clients can invoke through natural language requests. The server returns structured data that the AI client interprets and presents in a readable format.
Before starting, install and configure Gcore MCP Server with GCORE_TOOLS="ai,ai_ml,gpu_baremetal,gpu_virtual" to load AI/ML tools.

​
How it works

When a request is made in natural language, the AI client resolves project and region context, calls the appropriate MCP tool, and formats the JSON response for readability. Example request: 
List my GPU instances
The MCP server returns structured data for GPU-equipped instances: 
{
  "id": "38a68613-b89d-4ff5-9a86-940455a49919",
  "name": "ml-training-01",
  "status": "ACTIVE",
  "flavor": {
    "flavor_name": "g2-gpu-16-64-a2-2",
    "vcpus": 16,
    "ram": 65536
  },
  "metadata": {
    "os_distro": "ubuntu",
    "os_version": "22.04",
    "image_name": "ubuntu-22.04-x64"
  },
  "region": "Luxembourg-2"
}
The AI client presents this as readable text: 
Found 2 GPU instances in Luxembourg-2:

1. ml-training-01
   Status: ACTIVE
   Flavor: g2-gpu-16-64-a2-2 (16 vCPU, 64 GB RAM)
   GPU: 2x NVIDIA A2
   OS: Ubuntu 22.04

2. render-workstation
   Status: SHUTOFF
   Flavor: g2w-gpu-8-32-a2-1 (8 vCPU, 32 GB RAM)
   GPU: 1x NVIDIA A2
   OS: Windows Server 2022

​
GPU instances

GPU instances provide dedicated graphics processing units for AI/ML training, inference, rendering, and high-performance computing workloads.

​
List GPU instances

List my GPU instances
The response shows all instances with GPU flavors, including GPU type, vCPU count, RAM, and current status. 
Show GPU flavors available in Luxembourg
Response: 
GPU flavors in Luxembourg-2:

Linux flavors:
  g2-gpu-8-32-a2-1    8 vCPU, 32 GB RAM, 1x NVIDIA A2
  g2-gpu-16-64-a2-2   16 vCPU, 64 GB RAM, 2x NVIDIA A2
  g2-gpu-32-128-a2-4  32 vCPU, 128 GB RAM, 4x NVIDIA A2
  
Windows flavors:
  g2w-gpu-8-32-a2-1   8 vCPU, 32 GB RAM, 1x NVIDIA A2

High-performance (H100):
  Available in select regions. Contact support for access.

​
Create a GPU instance

Creating a GPU instance allocates dedicated GPU resources. 
Create a GPU instance with 8 vCPUs, 32 GB RAM, and NVIDIA A2 running Ubuntu 22.04
Response: 
Creating GPU instance...

Instance created:
  ID: 9a8b7c6d-5e4f-3a2b-1c0d-9e8f7a6b5c4d
  Name: gpu-instance-1
  Status: BUILD
  
Configuration:
  Flavor: g2-gpu-8-32-a2-1
  GPU: 1x NVIDIA A2
  vCPU: 8
  RAM: 32 GB
  OS: Ubuntu 22.04

The instance will be ready in 2-3 minutes.
GPU drivers are pre-installed on supported images.
For Windows with GPU: 
Create a Windows Server 2022 instance with GPU for rendering
Response: 
Creating Windows GPU instance...

Instance created:
  ID: 1b2c3d4e-5f6a-7b8c-9d0e-1f2a3b4c5d6e
  Name: win-gpu-1
  Status: BUILD
  
Configuration:
  Flavor: g2w-gpu-8-32-a2-1
  GPU: 1x NVIDIA A2
  vCPU: 8
  RAM: 32 GB
  OS: Windows Server 2022

Connect via RDP once the instance is ACTIVE.
GPU flavors have limited regional availability. The AI client suggests alternative regions when a requested GPU configuration is unavailable locally.

​
Manage GPU instance lifecycle

GPU instances support the same lifecycle operations as standard instances. 
Stop instance ml-training-01 to save costs
Response: 
Stopping instance ml-training-01...

Instance stopped.
Status: SHUTOFF
GPU resources released.

Billing for compute pauses while stopped.
Storage charges continue.

To restart: "Start instance ml-training-01"

Resize GPU instance ml-training-01 to 32 vCPUs and 128 GB RAM with 4 A2 GPUs
Response: 
Resizing instance ml-training-01...

This operation requires:
1. Stop the instance (will be done automatically)
2. Change flavor to g2-gpu-32-128-a2-4
3. Restart the instance

Proceed? (The AI waits for confirmation or proceeds based on context)

Resize completed:
  Previous: g2-gpu-16-64-a2-2 (2x A2)
  New: g2-gpu-32-128-a2-4 (4x A2)
  Status: ACTIVE

​
GPU clusters

GPU clusters provide multi-node GPU infrastructure for distributed training and large-scale inference. Available in Bare Metal and Virtual configurations.

​
List GPU clusters

List my GPU clusters
If no clusters exist: 
No GPU clusters found in Luxembourg-2.

To create a GPU Bare Metal cluster:
"Create a GPU Bare Metal cluster with 4 H100 nodes"

To create a GPU Virtual cluster:
"Create a GPU Virtual cluster with 2 A100 nodes"

​
Create a GPU Bare Metal cluster

GPU Bare Metal clusters provide dedicated physical servers with direct GPU access. 
Create a GPU Bare Metal cluster named ml-training with 4 H100 nodes
Response: 
Creating GPU Bare Metal cluster ml-training...

Cluster configuration:
  Name: ml-training
  Nodes: 4
  GPU per node: 8x NVIDIA H100 80GB
  Interconnect: 3.2 Tbit/s InfiniBand
  RAM per node: 2TB
  Storage per node: 8x 3.84TB NVMe

Estimated provisioning time: 15-30 minutes

Cluster creation initiated.
ID: 2c3d4e5f-6a7b-8c9d-0e1f-2a3b4c5d6e7f

Status updates:
  - Allocating hardware...
GPU Bare Metal cluster creation involves physical server allocation and takes 15-30 minutes. The AI client monitors progress and reports when all nodes are ready.

​
Create a GPU Virtual cluster

GPU Virtual clusters offer flexibility with faster provisioning. 
Create a GPU Virtual cluster named inference-dev with 2 A100 nodes
Response: 
Creating GPU Virtual cluster inference-dev...

Cluster configuration:
  Name: inference-dev
  Nodes: 2
  GPU per node: 8x NVIDIA A100 80GB
  Interconnect: Ethernet

Cluster creation initiated.
ID: 3d4e5f6a-7b8c-9d0e-1f2a-3b4c5d6e7f8a

Estimated time: 5-10 minutes

​
Manage cluster nodes

Once a cluster is created, individual nodes can be managed. 
Get console access for node 1 in cluster ml-training
Response: 
Console access for ml-training node 1:

URL: https://console.gcore.com/vnc/...
Token: abc123xyz
Valid for: 10 minutes

Open the URL in a browser to access the node console.

Reboot all nodes in cluster ml-training
Response: 
Rebooting all nodes in cluster ml-training...

Node 1: rebooting
Node 2: rebooting
Node 3: rebooting
Node 4: rebooting

All nodes will be back online in approximately 5 minutes.

​
Scale clusters

Add 2 more nodes to cluster ml-training
Response: 
Scaling cluster ml-training...

Current nodes: 4
Adding: 2
New total: 6

Provisioning additional nodes...
Estimated time: 15-20 minutes

​
AI inference (Everywhere Inference)

Everywhere Inference deploys trained AI models on edge inference nodes with automatic routing to the nearest endpoint.

​
List inference deployments

List my inference deployments
If no deployments exist: 
No inference deployments found.

Available actions:
  - "Deploy Llama 2 70B model" (from catalog)
  - "Deploy custom model from registry my-registry/model:v1"

​
Deploy a model from catalog

Deploy Llama 2 70B with 3 replicas in Europe
Response: 
Deploying Llama 2 70B...

Configuration:
  Model: meta-llama/Llama-2-70b-chat-hf
  Replicas: 3
  Regions: Luxembourg, Frankfurt, Amsterdam
  GPU type: L40S (auto-selected for model size)

Deployment initiated.
ID: 4e5f6a7b-8c9d-0e1f-2a3b-4c5d6e7f8a9b

Status: pulling model weights (this may take 10-15 minutes)

Endpoint will be available at:
https://llama2-70b-xxx.inference.gcore.com

​
Deploy a custom model

Deploy my custom model from registry docker.io/myorg/sentiment:v2 in Luxembourg
Response: 
Deploying custom model sentiment:v2...

Configuration:
  Image: docker.io/myorg/sentiment:v2
  Replicas: 1
  Region: Luxembourg

Deployment initiated.
ID: 5f6a7b8c-9d0e-1f2a-3b4c-5d6e7f8a9b0c

Status: pulling container image

Endpoint will be available at:
https://sentiment-xxx.inference.gcore.com

​
Check deployment status

Show status of deployment llama2-70b
Response: 
Deployment: llama2-70b

Status: running
Replicas: 3/3 healthy

Regions:
  Luxembourg: 1 replica, healthy
  Frankfurt: 1 replica, healthy
  Amsterdam: 1 replica, healthy

Endpoint: https://llama2-70b-xxx.inference.gcore.com
Traffic: 1,250 requests/hour

GPU utilization:
  Luxembourg: 45%
  Frankfurt: 62%
  Amsterdam: 38%

​
Scale deployments

Scale deployment llama2-70b to 5 replicas
Response: 
Scaling deployment llama2-70b...

Current replicas: 3
Target replicas: 5
Adding: 2 replicas (auto-placed in high-traffic regions)

Scaling in progress...
New replicas will be ready in 5-10 minutes.

​
Troubleshooting

​
GPU flavor not available

Show GPU availability in all regions
Response: 
GPU availability:

Luxembourg-2:
  A2: available
  A100: limited
  H100: contact support

Frankfurt:
  A2: available
  A100: available
  H100: limited

Amsterdam:
  A2: available
  L40S: available

Contact support for H100 or dedicated GPU allocations.

​
Cluster creation fails

Common causes:
  • Insufficient quota for GPU resources
  • Region capacity constraints
Show my GPU quota
Response: 
GPU quota for project default:

A2 GPUs: 4 used / 8 available
A100 GPUs: 0 used / 0 available (contact support to enable)
H100 GPUs: 0 used / 0 available (contact support to enable)

To request quota increase: support@gcore.com

​
Inference deployment not responding

Show logs for deployment sentiment-v2
Response: 
Recent logs for sentiment-v2:

[2026-04-15 10:30:15] Container started
[2026-04-15 10:30:20] Loading model weights...
[2026-04-15 10:30:45] Model loaded successfully
[2026-04-15 10:31:00] Listening on port 8080
[2026-04-15 10:35:22] ERROR: CUDA out of memory

Issue: Model requires more GPU memory than allocated.
Solution: Upgrade to larger GPU or reduce batch size.
For detailed GPU Cloud configuration, see About GPU Cloud. Inference deployment is documented in Deploy an AI model. GPU instance types are covered in Types of Virtual Machines.