Thunder Compute MCP Server
io.github.Thunder-Compute/thunder-compute
GPU cloud platform CLI for provisioning, managing, and monitoring AI/ML compute instances with SSH access and file transfer.
What is the Thunder Compute MCP server?
The Thunder Compute MCP server is an interface to Thunder Compute's GPU cloud platform, enabling AI agents to provision, manage, and monitor GPU instances, SSH keys, snapshots, and billing through the `tnr` CLI. It supports instance creation and deletion, secure SSH access, file transfers via SCP, port forwarding, and resource configuration for AI/ML workloads.
Thunder Compute is a high-performance GPU cloud platform optimized for AI/ML prototyping and experimentation. This MCP server exposes Thunder Compute's capabilities through an AI-accessible interface, allowing you to programmatically create and manage GPU instances, transfer files, establish secure connections, and monitor compute resources—all designed for fast provisioning and low-cost GPU access.
How to install Thunder Compute
Copy-paste configuration for popular MCP clients.
Tools & capabilities
Tools this server exposes to the agent.
Instance Management— Create, delete, and manage GPU instances with configurable compute resourcesSSH Access— Secure SSH session management and connection to remote instancesFile Transfer— SCP-based file upload and download between local and remote instancesPort Forwarding— Forward remote service ports to local accessStatus Monitoring— View instance status and resource utilizationAuthentication— Browser-based OAuth login or API token-based authenticationSSH Key Management— Configure and manage SSH keys for instance access
Use cases
- Provision GPU instances on-demand for training machine learning models
- Transfer training data and model checkpoints to/from GPU instances via SCP
- Monitor instance status and resource usage during long-running AI experiments
- Automate GPU instance lifecycle (create, configure, delete) for batch ML jobs
- Establish secure SSH tunnels to access Jupyter notebooks or other services running on GPU instances
Thunder Compute MCP server FAQ
It's an MCP interface to Thunder Compute's GPU cloud platform, allowing AI agents to create, manage, and monitor GPU instances, handle SSH access, transfer files, and manage billing—all optimized for AI/ML workloads.
Thunder Compute is a paid GPU cloud platform. Pricing details are available at thundercompute.com; the MCP server itself is open-source under the MIT License.
Use `tnr login` for browser-based OAuth authentication, or `tnr login --token <api_token>` if you've created a named API token in Console Settings. The MCP server uses these credentials to access your account.
The `tnr` CLI is available for macOS (Intel and Apple Silicon), Linux (via .tar.gz, .deb, .rpm, .apk), and Windows (via .msi, Scoop, or Winget).
Yes, the server supports SCP-based file transfers in both directions: upload files to instances with `tnr scp myfile.py 0:/path/` and download results with `tnr scp 0:/path/results.txt ./`.
You need a Thunder Compute account and valid API credentials. The MCP server connects remotely via https://api.thundercompute.com:8443/mcp and requires authentication before you can manage instances.
README (reference)
Source of truth, from the repository.
Overview
tnr is the official command-line interface for Thunder Compute, a high-performance cloud GPU platform built for AI/ML prototyping and experimentation.
Using a proprietary orchestration engine, Thunder Compute delivers fast provisioning, low-latency execution, and low-cost GPU compute designed for developer workflows.
The tnr CLI supports:
- Provisioning and managing GPU instances
- Configuring compute resources and instance specifications
- Secure SSH access and session management
- File transfer capabilities (SCP upload and download)
- Port forwarding for accessing remote services locally
- Automated update checks and version management
- Cross-platform support for macOS, Linux, and Windows
Documentation
- CLI Reference: https://www.thundercompute.com/docs/cli-reference
- API Reference: https://www.thundercompute.com/docs/api-reference
- Get Started: https://www.thundercompute.com/docs/quickstart
- Troubleshooting: https://www.thundercompute.com/docs/troubleshooting
Installation
Install tnr using one of the supported methods below. You may also download installers and binaries directly from the Latest Release page.
macOS
Download the .pkg installer for your platform from the Latest Release page (tnr_*_darwin_arm64.pkg for Apple Silicon, tnr_*_darwin_amd64.pkg for Intel).
Homebrew:
brew tap Thunder-Compute/tnr
brew install tnr
Linux
Download .tar.gz, .deb, .rpm, or .apk packages from the Latest Release page.
Install script (recommended):
curl -fsSL https://raw.githubusercontent.com/Thunder-Compute/thunder-cli/main/scripts/install.sh | bash
Windows
Download the .msi installer for your platform from the Latest Release page.
Scoop:
scoop bucket add tnr https://github.com/Thunder-Compute/scoop-bucket
scoop install tnr
Winget:
winget install Thunder.tnr
Build from Source
git clone https://github.com/Thunder-Compute/thunder-cli.git
cd thunder-cli
go build -o tnr
./tnr
Quick Start
tnr login # Authenticate with Thunder Compute
tnr create # Create a GPU instance
tnr status # View instance status
tnr connect 0 # Connect to your instance
# File transfers
tnr scp myfile.py 0:/home/ubuntu/
tnr scp 0:/home/ubuntu/results.txt ./
tnr delete 0 # Delete instance
tnr login opens a browser and returns to a loopback callback on
127.0.0.1. On a headless or remote machine where that callback cannot be
reached, create a named API token in Console Settings and run
tnr login --token <named_api_token> instead.
Local and staging testing (monorepo)
For a local API, run make cli-local ARGS="login", make cli-local ARGS="status", or make cli-local ARGS="logout --revoke" from the Thundernetes repository root. This builds cli/dist/tnr-local and targets http://localhost:8443, storing credentials in cli/dist/local-home separately from staging and production. Start your local API first; the runner does not start it. Browser login also needs the API's OAuth configuration and a running authorization page, normally the local Console at http://localhost:3000. For arguments requiring shell quoting, use bash cli/scripts/tnr-local.sh <arguments>.
From the Thundernetes repository root, run make cli-staging ARGS="login", make cli-staging ARGS="status", or make cli-staging ARGS="logout --revoke". Each invocation builds the current branch through Bazel and runs cli/dist/tnr-staging against the URL supplied through TNR_STAGING_API_URL. The monorepo Makefile supplies the default; the CLI scripts contain no staging endpoint. Credentials are saved separately in cli/dist/staging-home, an inherited TNR_API_TOKEN is ignored, and self-updates are disabled. These generated files are gitignored. For arguments requiring shell quoting, export TNR_STAGING_API_URL and run bash cli/scripts/tnr-staging.sh <arguments> directly.
Browser login requires the staging API's /v1/auth/cli/oauth-config endpoint and staging Connected App configuration. A 404 indicates the route is unavailable; a 503 can indicate missing OAuth configuration. Local builds do not deploy or configure the API.
License
This project is licensed under the MIT License. See LICENSE for details.
Related MCP servers

ContextBridge
MCP memory server for AI coding assistants. Windows Service, ONNX embeddings, SQLite vector search.
MCP-native web evidence and claim verification: cited, source-grounded evidence for AI agents.

Weather
US weather for AI agents: active NWS alerts by state, 5-period forecasts by lat/lon. Paid per call.
Shared AI + user PowerShell session on Windows - execute commands, run scripts, automate your PC
Complete SSH terminal for Linux with AI/user control and dual-stream visibility

TickDB Market Data
One MCP connection for real-time and historical forex, stock, crypto, and metals market data.