Massed Compute MCP Server
Enables AI assistants to manage Massed Compute GPU instances, including browsing inventory, launching and managing VMs, and auditing billing.
README
Massed Compute MCP Server
<!-- mcp-name: io.github.Massed-Compute/mcp -->
A Model Context Protocol server that lets AI assistants — Claude, Cursor, Codex, ChatGPT, and other MCP-compatible clients — interact with your Massed Compute account: browse GPU inventory, launch and manage VMs, audit billing.
Contents
Overview
The server exposes 14 tools that map 1:1 to documented /api/v1/* endpoints — no internal services, no undocumented calls. Issue a read-only key for analysis-only assistants and destructive tools (launch, restart, terminate, SSH-key changes) are hidden from the catalog entirely. Works with Claude Code, Claude Desktop, Cursor, and Codex.
| Tool | Required key | Returns |
|---|---|---|
gpu_inventory_list |
read-only | GPU configurations, pricing, regional capacity |
images_list |
read-only | VM image catalog |
instances_list |
read-only | Your running VM instances |
instances_get |
read-only | A single instance by UUID |
instances_launch |
full | Newly-launched instance details (incurs cost) |
instances_restart |
full | Restart confirmation |
instances_terminate |
full | Termination confirmation (destructive) |
coupon_information |
read-only | Coupon discount terms |
coupon_accepted_products |
read-only | Products a coupon applies to |
account_token_validation |
read-only | Token validity status |
account_billing |
read-only | Billing settings, recharge configuration |
ssh_keys_list |
read-only | Your SSH keys |
ssh_keys_create |
full | Newly-created key details |
ssh_keys_delete |
full | Deletion confirmation (destructive) |
Beyond raw tools, Massed Compute publishes Agent Skills — markdown workflow templates for common operations like GPU selection and cost auditing. Full docs at vm-docs.massedcompute.com/docs/category/mcp.
Installation
Step 1. Get your API key
Open vm.massedcompute.com/settings/api, create a key (read-only for analysis-only assistants; full-access to allow launch / restart / terminate / SSH-key changes), copy it.
Step 2. Pick an install path
- Hosted endpoint — point your client at the streamable-HTTP URL, zero install
- Local CLI — runs on your machine, key stored in your OS config dir
Hosted endpoint
Same 14 tools, same API key, nothing to install. Pick the snippet for your client:
Claude Code
claude mcp add --transport http massed-compute \
https://vm.massedcompute.com/api/mcp \
--header "Authorization: Bearer MC_TOKEN"
Cursor — ~/.cursor/mcp.json
{
"mcpServers": {
"massed-compute": {
"url": "https://vm.massedcompute.com/api/mcp",
"headers": { "Authorization": "Bearer MC_TOKEN" }
}
}
}
Codex — ~/.codex/config.toml
[mcp_servers.massed-compute]
url = "https://vm.massedcompute.com/api/mcp"
bearer_token_env_var = "MC_TOKEN"
enabled = true
Claude Desktop
Claude Desktop does not yet speak streamable-HTTP MCP, so use mcp-remote as a stdio↔HTTP bridge. Config at ~/Library/Application Support/Claude/claude_desktop_config.json (macOS) or %APPDATA%\Claude\claude_desktop_config.json (Windows):
{
"mcpServers": {
"massed-compute": {
"command": "npx",
"args": [
"-y", "mcp-remote",
"https://vm.massedcompute.com/api/mcp",
"--header", "Authorization: Bearer MC_TOKEN"
]
}
}
}
Local CLI
Install via whichever ecosystem you prefer:
npm install -g massed-compute-mcp # Node >= 20
# or
pip install massed-compute-mcp # Python >= 3.10
# or
uv tool install massed-compute-mcp # fast Python install via uv
Run one-shot setup:
massed-compute-mcp init
init prompts for the key, validates it upstream, stores it at 0600 (POSIX), detects installed MCP clients (Claude Code, Cursor, Claude Desktop, Codex), and offers to wire each one. A timestamped backup is taken before any client edit. Restart wired clients to pick up the tools.
Config file location:
- Linux:
$XDG_CONFIG_HOME/massed-compute/config.json(falls back to~/.config/...) - macOS:
~/Library/Application Support/massed-compute/config.json - Windows:
%APPDATA%\massed-compute\config.json
Non-interactive (CI / scripts):
# Key from env, auto-wire every detected client
MASSED_COMPUTE_API_KEY=<your-key> massed-compute-mcp init --yes
# Key from a file, wire only specific clients
massed-compute-mcp init --token-file ~/keys/mc --yes --clients claude-code,cursor
# Store the key only; don't touch any client config
massed-compute-mcp init --yes --no-install-clients
Add or remove a client later:
massed-compute-mcp install-client cursor # claude-desktop | codex | claude-code
massed-compute-mcp uninstall-client cursor
install-client is idempotent — re-running when the entry already matches is a silent no-op.
Verifying the connection
massed-compute-mcp doctor
doctor confirms the stored key still works, prints the tool catalog, and shows copy-pasteable snippets for any clients you didn't auto-wire. Or just ask your assistant "Validate my Massed Compute API key." — a { message: "Valid Token" } response confirms the wiring.
CLI reference
| Command | What it does |
|---|---|
massed-compute-mcp (no args) |
Run the MCP server over stdio |
massed-compute-mcp init |
First-run setup: prompt, validate, store, wire clients |
massed-compute-mcp doctor |
Verify the stored key and print client snippets |
massed-compute-mcp install-client <id> |
Wire a single client (claude-code, cursor, claude-desktop, codex) |
massed-compute-mcp uninstall-client <id> |
Remove our entry from a client config |
massed-compute-mcp config show |
Print resolved config path, masked key, resolution chain |
massed-compute-mcp logout |
Delete the stored API key |
massed-compute-mcp tools [--json] |
Print the tool catalog (no upstream call) |
massed-compute-mcp version |
Print the version |
Key resolution
When the server starts, the API key is taken from the first source that provides one:
--token <value>CLI flag--token-file <path>CLI flag (first line, trimmed)MASSED_COMPUTE_API_KEYenvironment variable- Stored config file written by
init
If none of those are set, the server exits non-zero with a pointer to massed-compute-mcp init. Override the upstream with MASSED_COMPUTE_API_BASE_URL (default https://vm.massedcompute.com).
Resources
- Full MCP documentation
- Massed Compute API reference
- Massed Compute console
- Model Context Protocol spec
License
MIT — see LICENSE.
推荐服务器
Baidu Map
百度地图核心API现已全面兼容MCP协议,是国内首家兼容MCP协议的地图服务商。
Playwright MCP Server
一个模型上下文协议服务器,它使大型语言模型能够通过结构化的可访问性快照与网页进行交互,而无需视觉模型或屏幕截图。
Magic Component Platform (MCP)
一个由人工智能驱动的工具,可以从自然语言描述生成现代化的用户界面组件,并与流行的集成开发环境(IDE)集成,从而简化用户界面开发流程。
Audiense Insights MCP Server
通过模型上下文协议启用与 Audiense Insights 账户的交互,从而促进营销洞察和受众数据的提取和分析,包括人口统计信息、行为和影响者互动。
VeyraX
一个单一的 MCP 工具,连接你所有喜爱的工具:Gmail、日历以及其他 40 多个工具。
graphlit-mcp-server
模型上下文协议 (MCP) 服务器实现了 MCP 客户端与 Graphlit 服务之间的集成。 除了网络爬取之外,还可以将任何内容(从 Slack 到 Gmail 再到播客订阅源)导入到 Graphlit 项目中,然后从 MCP 客户端检索相关内容。
Kagi MCP Server
一个 MCP 服务器,集成了 Kagi 搜索功能和 Claude AI,使 Claude 能够在回答需要最新信息的问题时执行实时网络搜索。
e2b-mcp-server
使用 MCP 通过 e2b 运行代码。
Neon MCP Server
用于与 Neon 管理 API 和数据库交互的 MCP 服务器
Exa MCP Server
模型上下文协议(MCP)服务器允许像 Claude 这样的 AI 助手使用 Exa AI 搜索 API 进行网络搜索。这种设置允许 AI 模型以安全和受控的方式获取实时的网络信息。