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., "@vastai-mcp find me a cheap GPU with at least 24GB VRAM"
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.
Vast.ai GPU cloud marketplace and instance management MCP server.
Wraps the Vast.ai REST API (https://console.vast.ai/api/v0) so an LLM agent can search the GPU marketplace, rent machines, manage lifecycle, run commands, and inspect billing.
Related MCP server: quads-mcp
Tool | Parameters | Description |
|
| Search the GPU marketplace (POST /bundles/) |
|
| GPU benchmark data (GET /benchmarks/) |
|
| Search user and public templates (GET /template/) |
|
| Current supply/demand/pricing snapshot |
|
| Historical price/supply time-series |
|
| List user's instances (GET /instances/) |
|
| Instance details (GET /instances/{id}/) |
|
| Rent a GPU (PUT /asks/{offer_id}/) |
|
| start/stop/label/reboot/recycle |
|
| Permanently destroy (DELETE /instances/{id}/) |
|
| Change spot bid price |
|
| Request instance logs |
|
| Run a remote command |
| none | Current account + credit balance |
|
| Escape hatch for any other endpoint |
|
| Register a local public key on the Vast.ai account (POST /ssh/) |
| none | List registered SSH keys (GET /ssh/) |
|
| Resolve |
|
| Run a command over SSH (paramiko) |
|
| SFTP upload/download between local disk and instance |
The query body for search_offers follows the Vast.ai filter operator convention: each field maps to {eq, neq, gt, lt, gte, lte, in, notin}. Example:
{
"gpu_name": {"in": ["RTX 4090"]},
"num_gpus": {"gte": 1},
"reliability": {"gte": 0.99},
"verified": {"eq": true},
"rentable": {"eq": true},
"type": "ondemand",
"limit": 5
}Variable | Required | Description |
| yes | Vast.ai API key (https://cloud.vast.ai/manage-keys/) |
| no | Default path to your SSH private key, used by |
Only requirement: uv (https://docs.astral.sh/uv/). Nothing else to install.
claude mcp add vastai-mcp -e VASTAI_MCP_API_KEY=<value> -- uvx --from git+https://github.com/dam2452/vastai-mcp.git vastai-mcp{
"mcpServers": {
"vastai-mcp": {
"command": "uvx",
"args": ["--from", "git+https://github.com/dam2452/vastai-mcp.git", "vastai-mcp"],
"env": { "VASTAI_MCP_API_KEY": "<value>" }
}
}
}After pushing a new version: uv cache clean and restart the client.
uv run --directory . vastai-mcpTests (manual):
uv run --directory . --with pytest pytest test/Change the bid price of an interruptible instance (PUT /instances/bid_price/{id}/).
Examples: change_bid(instance_id=12345, price=0.25)
Inputs: price (required), instance_id (required). Returns: object
Rent a GPU by accepting offer_id (PUT /asks/{offer_id}/).
Either image or template_hash_id must be provided. Use price to create
an interruptible (spot) instance; omit it for on-demand.
env is a flat dict: regular vars as key/value pairs, port mappings as
{"-p 8000:8000": "1"}. Use onstart to launch your app on SSH/Jupyter
runtypes (their entrypoint is replaced by Vast's).
Examples: create_instance(offer_id=12345678, image="ubuntu:22.04", disk=16, runtype="ssh_direct") create_instance(offer_id=12345678, image="vllm/vllm-openai:latest", disk=50, runtype="ssh_direct", env={"MODEL_ID": "deepseek-ai/DeepSeek-R1-Distill-Llama-8B", "-p 8000:8000": "1"}, onstart="vllm serve $MODEL_ID --port 8000") create_instance(offer_id=12345678, template_hash_id="4e17788f74f075dd9aab7d0d4427968f", disk=100) create_instance(offer_id=12345678, image="ubuntu:22.04", disk=16, runtype="ssh_direct", price=0.20)
Inputs: offer_id (required), env, disk, image, +10 more. Returns: object
Permanently destroy an instance and all its data (DELETE /instances/{id}/).
Irreversible. Call carefully.
Examples: destroy_instance(instance_id=12345)
Inputs: instance_id (required). Returns: object
Execute a constrained remote command on a STOPPED instance (PUT /instances/command/{id}/).
Only works on instances whose actual_status is stopped (transient launch
commands like downloading an image on first start). For running instances,
the API rejects this with: "Execute command only avail on stopped instances.
Use ssh to run commands on running instances." — use ssh_exec instead.
The API additionally enforces a command whitelist — unknown commands return
400 "Invalid command given". ls-style inspection commands are accepted;
nvidia-smi, whoami, arbitrary binaries are not. For arbitrary commands
on a running instance, use ssh_exec.
Examples: execute_command(instance_id=12345, command="ls /workspace")
Inputs: command (required), instance_id (required). Returns: object
Retrieve the authenticated user's account info and credit balance (GET /users/current/).
Examples: get_current_user()
Inputs: no named parameters. Returns: object
Current snapshot of supply, demand, and pricing across GPU types (GET /metrics/gpu/current/).
Examples: get_gpu_metrics() get_gpu_metrics(gpu_name=["RTX 4090", "A100"], verified=True)
Inputs: gpu_name, verified, datacenter. Returns: object
Time-series of GPU supply/demand/pricing (GET /metrics/gpu/history/).
Returns ~20 sampled data points over the window by default.
Examples: get_gpu_trends() get_gpu_trends(gpu_names=["RTX 4090"], window_hours=168)
Inputs: step, gpu_names, window_hours. Returns: object
Retrieve details of a specific instance (GET /instances/{id}/).
Examples: get_instance(instance_id=12345)
Inputs: instance_id (required). Returns: object
Resolve an instance's direct SSH endpoint (host, port) and return a
ready-to-run ssh command. Requires actual_status == "running" and a
ssh_direct or jupyter_direct runtype.
Examples: get_instance_ssh_info(instance_id=12345)
Inputs: instance_id (required). Returns: object
List the authenticated user's instances (GET /instances/).
Examples: list_instances() list_instances(label="vllm-inference")
Inputs: label. Returns: object
List SSH keys registered on the Vast.ai account (GET /ssh/).
Examples: list_ssh_keys()
Inputs: no named parameters. Returns: object
Control instance state: start/stop/label/reboot/recycle.
start/stop/label: PUT /instances/{id}/ with state or label.
reboot: PUT /instances/reboot/{id}/ (stops and starts, preserves GPU priority).
recycle: PUT /instances/recycle/{id}/ (destroys and recreates from fresh image pull).
For action="label", the label argument is required.
Examples: manage_instance(instance_id=12345, action="stop") manage_instance(instance_id=12345, action="start") manage_instance(instance_id=12345, action="label", label="my-server") manage_instance(instance_id=12345, action="reboot")
Inputs: action (required), instance_id (required), label. Returns: object
Escape hatch for any Vast.ai REST endpoint not covered by other tools.
path is appended to https://console.vast.ai/api/v0 (include the leading
slash, e.g. "/volumes/").
Examples: raw_request(method="GET", path="/volumes/") raw_request(method="POST", path="/ssh/", body={"ssh_key": "ssh-rsa AAA..."})
Inputs: path (required), method (required), body, params. Returns: object
Register a local SSH public key on the Vast.ai account (POST /ssh/).
The key is applied automatically to all current instances and any new instances created afterwards. Reads the .pub file from disk — never uploads your private key.
Examples: register_ssh_key(name="laptop", public_key_path="C:/Users/me/.ssh/id_ed25519.pub") register_ssh_key(name="ci", public_key_path="/home/runner/.ssh/id_ed25519.pub")
Inputs: name (required), public_key_path (required). Returns: object
Retrieve all GPU benchmark records (GET /benchmarks/).
Returns a list of benchmark entries with gpu_name, value, type, etc.
Examples: search_benchmarks()
Inputs: no named parameters. Returns: object
Search the Vast.ai GPU marketplace for rentable machine offers.
The body is sent as-is to POST /bundles/. Top-level keys commonly include
type ("ondemand" | "bid"), limit, and filter objects. Each filter maps
a field name to an operator object: {eq, neq, gt, lt, gte, lte, in, notin}.
Examples: search_offers({"gpu_name": {"in": ["RTX 4090"]}, "num_gpus": {"gte": 1}, "reliability": {"gte": 0.99}, "verified": {"eq": True}, "rentable": {"eq": True}, "type": "ondemand", "limit": 5}) search_offers({"type": "bid", "gpu_ram": {"gte": 24000}, "order": [["dph_total", "asc"]], "limit": 10})
Inputs: query (required). Returns: object
Search user and public templates (GET /template/).
The endpoint always requires a select_filters body (pass {} for "all").
This tool defaults to {} when called without arguments.
Examples: search_templates() search_templates(select_filters={"tag": {"eq": "pytorch"}})
Inputs: select_filters. Returns: object
Request instance logs (PUT /instances/request_logs/{id}). Logs are uploaded to S3; the response contains a URL to fetch them.
Examples: show_instance_logs(instance_id=12345) show_instance_logs(instance_id=12345, tail=5000)
Inputs: instance_id (required), tail. Returns: object
Execute a command on a running instance over SSH (paramiko).
Resolves the instance's ssh_host/ssh_port automatically. The private
key stays local — never sent to the Vast.ai API.
key_path defaults to the VASTAI_MCP_SSH_KEY_PATH env var if set.
Examples: ssh_exec(instance_id=12345, command="nvidia-smi") ssh_exec(instance_id=12345, command="ls /workspace", key_path="/home/me/.ssh/id_ed25519") ssh_exec(instance_id=12345, command="python train.py", user="root", timeout=300)
Inputs: command (required), instance_id (required), user, timeout, +1 more. Returns: object
Transfer a single file between the local machine and an instance over SFTP.
upload: local file -> instance path
download: instance path -> local file
Directories are not supported — call once per file. Resolves the instance's
ssh_host/ssh_port automatically. The private key stays local.
key_path defaults to the VASTAI_MCP_SSH_KEY_PATH env var if set.
Examples: ssh_transfer(direction="upload", instance_id=12345, local_path="./train.py", remote_path="/workspace/train.py") ssh_transfer(direction="download", instance_id=12345, local_path="./results.tar", remote_path="/workspace/results.tar")
Inputs: direction (required), local_path (required), instance_id (required), remote_path (required), +3 more. Returns: object
Recent tool additions, removals, and schema changes observed during successful MCP inspections.
Loading tool schema changelog…
Scored across 20 tools
Each tool targets a distinct resource and action: search_offers/search_benchmarks/search_templates operate on different datasets, get_gpu_metrics vs get_gpu_trends differ by snapshot vs history, and instance tools are clearly separated (list/get/create/manage/destroy). Even potentially similar tools like execute_command and ssh_exec are explicitly differentiated by stopped vs running instance state.
All tools follow a consistent verb_noun snake_case pattern (search_offers, list_instances, create_instance, destroy_instance, ssh_exec, ssh_transfer). The naming is predictable and consistent, with no camelCase or mixed conventions.
20 tools is at the upper end of 'well-scoped' but appropriate for a full Vast.ai API surface covering marketplace search, instance CRUD, SSH operations, metrics, and user info. The count feels justified given the breadth of features, though slightly heavy compared to a typical 3-15 tool server.
The surface covers the main GPU rental workflow: search offers, create/read/delete instances, manage state, run commands, transfer files, and fetch metrics. Minor gaps like volume management exist, but the raw_request escape hatch covers any missing endpoint, preventing dead ends.
An MCP server that provides an API to LLMs to manage their JumpCloud resources.
Hosted MCP server for live public-data APIs and Skills for AI agents.
MCP server exposing the Backtest360 engine API as tools for AI agents.
Nifty's MCP server — exposes tasks, projects, messages, and files as tools for AI agents.
