ComputeNexa MCP Server
OfficialClick on "Deploy Server".
Wait a few minutes for the server to deploy. Once ready, it will show a "Started" state.
In the chat, type
@followed by the MCP server name and your instructions, e.g., "@ComputeNexa MCP Server Find the cheapest 48GB GPU in stock and launch it for two hours."
That's it! The server will respond to your query, and you can continue using it as needed.
Here is a step-by-step guide with screenshots.
ComputeNexa MCP server
Let your AI assistant find, launch and stop private GPUs on ComputeNexa. Ask it "Which GPUs with 48 GB or more are in stock?", "Launch the cheapest one with Data Science for two hours" or "Stop it": it uses your team's API key, shows you what each launch costs, and only spends money after you agree.
Works with Claude Desktop, Claude Code, Cursor, VS Code and other assistants that support the Model Context Protocol. It runs on your computer and talks to the ComputeNexa API; your key stays in your own settings.
Set up
In the ComputeNexa dashboard, open API keys and create a key:
Look only to browse GPUs, machines and spending;
Launch and stop to also launch, extend and stop machines.
Install uv (it runs the server with
uvx, no separate install needed).Add the server to your assistant (below), with your key as
CNX_API_KEY.
Claude Code
claude mcp add --env CNX_API_KEY=cnxk_... --transport stdio computenexa -- uvx computenexa-mcpAdd --scope user to use it in every project.
Claude Desktop
Settings → Developer → Edit config (claude_desktop_config.json):
{
"mcpServers": {
"computenexa": {
"command": "uvx",
"args": ["computenexa-mcp"],
"env": { "CNX_API_KEY": "cnxk_..." }
}
}
}Cursor
~/.cursor/mcp.json (every project) or .cursor/mcp.json (one project): the same mcpServers entry as Claude
Desktop.
VS Code
.vscode/mcp.json; VS Code asks for the key once and keeps it out of the file:
{
"inputs": [
{ "type": "promptString", "id": "cnx-key", "description": "ComputeNexa API key", "password": true }
],
"servers": {
"computenexa": {
"type": "stdio",
"command": "uvx",
"args": ["computenexa-mcp"],
"env": { "CNX_API_KEY": "${input:cnx-key}" }
}
}
}Related MCP server: gpu.nz MCP server
What it can do
Tool | What it does | Key |
| Your email, team, balance, GPU limit and the key's permissions | any |
| GPUs with live prices and stock, cheapest first; filter by app, region, memory | any |
| The apps a GPU starts with, their models, and what's installed | any |
| Your machines; how to connect (SSH, browser apps, API or endpoint address) | any |
| Spending and GPU-hours; finished sessions and their cost | any |
| Saved projects; Endpoints and the machine each points at | any |
| Launch a GPU (quote first, then launch when you agree) | launch and stop |
| Add time (quote first) | launch and stop |
| Renew a machine automatically until stopped (quote first) | launch and stop |
| End a machine now; unused time is refunded | launch and stop |
| A stable HTTPS address for your own server; its key | launch and stop |
A look-only key only gets the read tools (the first six rows).
Safety
Nothing is charged without you.
launch_machine,extend_machineandset_keep_runningfirst return a quote, e.g. "NVIDIA RTX A6000 with the datascience app for 2 h: $1.44 charged now ($0.72/h); unused time is refunded if it stops early." Only a second call withconfirm=trueand the quoted total goes ahead; if the price went up in between, it's refused.A limit per call: a launch, extension or keep-running renewal costing more than $25 is refused. Change it with
CNX_MCP_MAX_LAUNCH_USD(in dollars) in the server'senv. Your team's GPU limit and prepaid balance still apply on top.Your API key never appears in a tool result. Endpoint keys are shown only when you create an endpoint or ask for its key.
Tools that spend money or end a machine are marked as such, so your assistant asks before using them.
Settings
Variable | |
| Required. A ComputeNexa API key ( |
| Most one call may spend, in US dollars. Default |
| Only for testing against another ComputeNexa API |
Develop
uv sync
uv run pytest
uv run ruff check src testsBuilt on the ComputeNexa Python SDK and the official MCP Python SDK.
Questions: support@computenexa.com · Security issues: see SECURITY.md.
Available Tools
15 toolscreate_endpointCreate an endpointA
A new Endpoint (owners and admins): a stable HTTPS address and key for your own server on port 8000, e.g. one per customer you serve. Launch machines for it with launch_machine(endpoint=...). Free. Its key is shown once here; get_endpoint_key shows it again.
Inputs: name (required). Returns: id, key, url, name, +1 more
extend_machineExtend a machineA
Add time to a machine at the price it launched with. Costs money: quote first (confirm=false), then confirm=true with expected_total_usd once the person agrees.
Inputs: hours (required), machine_id (required), confirm, expected_total_usd. Returns: quote, status, machine
get_endpoint_keyGet an endpoint's keyARead-onlyIdempotent
An Endpoint's key (owners and admins). Only when the person asks for it: it gives access to their server.
Inputs: endpoint (required). Returns: key, url, endpoint
get_machineGet a machineARead-onlyIdempotent
One machine: its state, how to connect (SSH command, browser apps, API or endpoint address), time left, keep-running status and costs. Never its keys.
Inputs: machine_id (required). Returns: id, app, gpu, label, +23 more
get_usageGet usageBRead-onlyIdempotent
Spending and GPU-hours over the last N days, by GPU, app, day and project.
Inputs: days. Returns: days, daily, by_app, by_gpu, +6 more
launch_machineLaunch a machineA
Launch a GPU machine. Costs money: call first with confirm=false to get a quote (nothing happens), show it to the person, and only after they agree call again with the same arguments, confirm=true and expected_total_usd. The booked time is charged up front from the team's balance.
Inputs: app (required), gpu (required), hours (required), label, +8 more. Returns: quote, status, machine
list_appsList appsARead-onlyIdempotent
The apps a GPU can start with (basic, datascience, chat, imagegen, modelapi, voice): what each is for, the
models it offers and which model we recommend per GPU memory. Give app for the software installed.
Inputs: app, package. Returns: apps, software
list_endpointsList endpointsARead-onlyIdempotent
Endpoints: stable HTTPS addresses for your own server on a machine's port 8000, and the machine each points at now. Never their keys.
Inputs: no named parameters. Returns: endpoints
list_gpusList GPUsARead-onlyIdempotent
GPUs with live prices (US dollars per GPU per hour) and stock, cheapest first.
Inputs: app, region, min_memory_gb, include_sold_out. Returns: gpus, region
list_machinesList machinesBRead-onlyIdempotent
Your machines, newest first: state, GPU, app, price, time left and cost so far.
Inputs: limit, whole_team, active_only. Returns: machines
list_projectsList projectsBRead-onlyIdempotent
Saved projects (files restored into ~/work on machines launched with them) and their size.
Inputs: no named parameters. Returns: projects
list_receiptsList receiptsBRead-onlyIdempotent
Finished sessions and what each cost in the end (charged minus refunded).
Inputs: days, limit. Returns: receipts, total_net_usd
set_keep_runningKeep a machine runningA
Keep running: the machine renews itself hours at a time from the team's balance, at its launch price,
until it's stopped. Turning it on spends money at every renewal: quote first, then confirm=true with
expected_renewal_usd. Turning it off (hours=null) needs no confirmation.
Inputs: hours (required), machine_id (required), confirm, expected_renewal_usd. Returns: quote, status, machine
stop_machineStop a machineADestructiveIdempotent
End a machine now: stopped if running, cancelled with a full refund if still starting. Files outside a project are deleted; a project is saved first. Unused time is refunded. Can't be undone.
Inputs: machine_id (required). Returns: machine, refunded_usd
whoamiWho am IARead-onlyIdempotent
Who this server's API key acts as: email, team and role, the team's balance and GPU limit, and whether the key may launch and stop machines.
Inputs: no named parameters. Returns: role, team, email, api_key, +4 more
Tool Schema Changelog
Recent tool additions, removals, and schema changes observed during successful MCP inspections.
Loading tool schema changelog…
TDQS
Scored across 15 tools
Each tool targets a distinct resource or lifecycle action: GPUs/apps discovery, machine launch/list/get/extend/stop/keep-running, endpoints, projects, usage, receipts, and identity. The potentially closest pair, extend_machine and set_keep_running, is clearly separated by one-off time extension versus automatic renewal, with both having explicit confirmation semantics.
Nearly all tools follow a snake_case verb_noun pattern such as list_machines, get_machine, launch_machine, stop_machine, create_endpoint, and get_endpoint_key. whoami is the lone conventional outlier, and set_keep_running is slightly more phrasal, but the naming remains predictable overall.
Fifteen tools fit the server's scope well: discovery, machine lifecycle, spending, projects, endpoints, and identity each contribute a small set of operations. No obvious redundant tools inflate the surface, and the set stays within a manageable range.
Core GPU machine workflows are well covered: discovering GPUs/apps, launching, inspecting, extending, stopping, and checking costs. However, projects expose only list_projects with no create/get/delete lifecycle, and endpoints expose create/list/get-key but no delete or revoke operation, leaving notable dead ends for resource management.
Maintenance
Related MCP Connectors
Rent cloud GPUs from AI agents: live prices, VRAM estimates, launch, logs and stop.
On-demand GPU nodes for agents: create nodes, run commands, and submit jobs, billed by the minute.
AI cloud for agents: launch GPUs, run jobs that stop their own billing, schedule runs.
Live cloud GPU prices, observed stock and 12-month price history across 30+ providers (H100, H200, A100, RTX 5090 and more), refreshed hourly. Read-only. Needs a GPU Finder API key sent as "Authorization: Bearer <key>"; free key with 60 data calls a month at https://gpufinder.dev/developers. Listing tools and get_usage are free.
Related MCP Servers
- AlicenseAqualityCmaintenanceEnables AI agents to retrieve live GPU rental prices, compare models, and access market summaries and historical price data through four read-only MCP tools.4MIT

gpu.nz MCP serverofficial
AlicenseNot gradedqualityBmaintenanceEnables AI agents to search live GPU offers, estimate model VRAM, rent and monitor cloud GPU servers with ready templates, read logs, and stop or destroy them when done. Browsing and VRAM estimates work without an API key, while server management requires one.1MIT
Runcrateofficial
AlicenseNot gradedqualityBmaintenanceEnables AI agents to work with a GPU cloud: check live GPU prices, launch GPUs on ready images, run commands as jobs that stop billing when the command exits, schedule runs, move files, and check spend. Access is scoped to one signed-in environment and capped by its monthly budget.MIT
GPU Finder MCP serverofficial
AlicenseAqualityBmaintenanceEnables agents to query live cloud GPU pricing, stock availability, and up to 12 months of price history across 30+ hyperscalers, neoclouds, and marketplaces, including comparisons of cheapest listed versus cheapest in-stock offers. All tools are read-only, with provider listings and GPU catalogs available at no cost.863 npmMIT