mnvitop-mcp
Click 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., "@mnvitop-mcp find me a free GPU on any host"
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.
mnvitop
One terminal for the GPUs of all your ssh hosts. It runs nvidia-smi over ssh
and shows every host's GPUs, their memory, utilisation history and running
processes in a single live view. Nothing needs to be installed on the remote
side beyond nvidia-smi and ps. It also ships an MCP server, so an agent can
ask where a free GPU is.

Install
uv tool install mnvitopYou need Python 3.11 or newer locally and ssh access to the hosts. To get the
MCP server as well, install mnvitop[mcp].
Related MCP server: Maestro MCP
Use
mnvitop h20 semi pro6000Each argument is an ssh destination, either an alias from ~/.ssh/config or
user@host. Ctrl-C quits.
mnvitop -a # temperature, process cpu and memory, more processes per GPU
mnvitop --once # print once and exit
mnvitop --json # one JSON line per round, for loggingmnvitop --help lists the rest. To stop typing the hosts every time, put them
in ~/.config/mnvitop/hosts.toml, following
hosts.example.toml.
For agents, mnvitop-mcp is a stdio MCP server with three tools: find_gpus,
fleet_status and list_hosts. Register it in Claude Code with
claude mcp add mnvitop -- mnvitop-mcp. The repository also carries a matching
skill in skills/gpu-fleet.
MIT license.
Available Tools
3 toolsfind_gpusA
Find where a job that needs count idle GPUs with at least min_free_gb free VRAM each can run right now.
Call this first for any placement question. Returns found, a summary line, placements best first (the least loaded host that has enough idle GPUs, with the GPU indices to use), and when nothing fits, alternatives: hosts with fewer idle GPUs and busy GPUs that still have the VRAM (shared with someone else's job, so only for small experiments).
Inputs: count, hosts, min_free_gb.
fleet_statusA
Live GPU status of every configured host (or just the named ones), probed now over ssh; takes 1-5 s.
Returns a fleet summary line, then per host: free and busy GPU counts, free VRAM, load, and per GPU the utilisation, VRAM used and free, temperature and (with processes=true) the largest processes with user, pid, VRAM, age and command. A GPU is busy when a process holds it or it has more than five percent of its memory taken. Hosts that do not answer come back with ok=false and an error instead of raising.
Inputs: hosts, processes.
list_hostsA
List the configured GPU hosts: display name and ssh target. No remote calls; use it to learn the names the other tools accept.
Inputs: no named parameters.
Tool Schema Changelog
Recent tool additions, removals, and schema changes observed during successful MCP inspections.
Loading tool schema changelog…
TDQS
Scored across 3 tools
Each tool has a clearly distinct purpose: listing configured hosts, querying live status, and finding placement for GPU jobs. No overlap or ambiguity between them.
Two tools use verb_noun pattern (list_hosts, find_gpus) while fleet_status is a noun phrase, creating a minor inconsistency. However, the names are short, clear, and follow a predictable theme of fleet management.
Three tools are ideal for a focused GPU fleet monitoring and placement server. Each tool serves a necessary and non-redundant function, and the count is well within the recommended range.
The server covers listing, status probing, and placement suggestion, which are core to its stated purpose. Minor gaps exist (e.g., no tool to modify host configuration or trigger actions), but for monitoring and placement discovery, it is complete.
Maintenance
Related MCP Connectors
Run GPU acceptance tests on your NVIDIA GPU nodes over SSH: DCGM, GEMM, HBM/PCIe bandwidth, NCCL.
On-demand GPU nodes for agents: create nodes, run commands, and submit jobs, billed by the minute.
Rent cloud GPUs from AI agents: live prices, VRAM estimates, launch, logs and stop.
Live GPU rental market: 2,500+ offers across a dozen provider feeds. History, watches, limit orders.
Related MCP Servers
- FlicenseNot gradedqualityAmaintenanceAn MCP server for monitoring and managing multi-cluster Slurm GPU jobs, enabling AI agents to execute commands, check allocations, and explore logs across HPC clusters.1-
- AlicenseNot gradedqualityDmaintenanceEnables AI agents to orchestrate a heterogeneous machine fleet via SSH, with unified command execution, file transfer, and dispatch of coding agents across platforms.2Apache 2.0
- FlicenseNot gradedqualityBmaintenanceEnables AI coding agents to run policy-constrained Python jobs on lab NVIDIA GPU hosts, including GPU inspection, device reservation, job launching and monitoring, and owner-scoped process stopping.2-
- FlicenseAqualityCmaintenanceEnables managing remote hosts from ~/.ssh/config over OpenSSH, including availability checks, command execution with sudo and timeouts, background jobs, SSH config and known_hosts management, and llms.txt-based documentation retrieval.19-