GPU cloud MCP servers compared
We checked official GPU cloud MCP servers from six providers. Only Lium's hosted server returned live GPU prices without an account in our checks. Lium also documents daily, monthly and lifetime USD budgets per agent key.
Lium team3 min read
comparison
mcp
ai-agents
claude-code
codex
An MCP server lets Claude Code, Codex or any MCP client find, rent and remove GPUs without a person clicking through a console. We checked six GPU clouds on 29 October. Three run an official MCP server that rents GPUs: Lium, Runpod and Thunder Compute. Vast.ai, Lambda and Modal had no official one.
Among the hosted servers we checked, Lium was the only one to return live GPU prices without an account. It also documents per-key USD spending budgets; Runpod's and Thunder Compute's MCP guides did not describe equivalent limits.
Lium's hosted MCP combines documentation, live GPU discovery and authenticated rental tools at one endpoint:
https://docs.lium.io/mcp.Why Lium
- Live prices before signup.
search_gpusandget_gpu_priceswork with no sign-in, so an agent can quote a job first. Runpod's and Thunder Compute's servers returned401without a token. - A budget per agent key. Each API key can have a daily, monthly and lifetime USD budget. A key at its budget, or without room for a new pod's first 15 minutes, can't start a pod (
402), and its running pods are removed when spending reaches the limit, so the limit doesn't depend on the agent behaving. - Rent by spec. Give
rent_poda GPU model and count and Lium picks the cheapest matching node. Anidempotency_keystops a retry from renting twice. - A read-only option.
/mcp/directoryissues a key that can list pods and read the balance but can't rent or spend. - Per-second billing.
remove_podstops billing when it's accepted.
How we checked
Last checked: 29 Oct 2026, 09:54 UTC. We read each cloud's MCP docs, searched the official MCP Registry and sent an unauthenticated MCP
initialize request to each hosted endpoint. We list only official servers run by the cloud itself. Third-party servers exist for most clouds, but they're not supported by the provider.The comparison
Runpod's server also manages Serverless endpoints, templates and network volumes. Thunder Compute's adds
run_command, snapshots and port forwarding. Lium's hosted server sticks to finding, renting and removing pods; volumes, backups and clusters are in the Lium API.In the MCP Registry, Lium is listed as
io.github.Datura-ai/lium-gpu-cloud (hosted) and io.github.Datura-ai/lium (local), and Thunder Compute as io.github.Thunder-Compute/thunder-compute. Runpod's only entry is a third-party one, com.mcparmory/runpod.What a price call returns
With no account,
get_gpu_prices for H100 returned H100 80GB HBM3 from $1.60 per GPU-hour, with 63 GPUs free on 8 nodes, at 09:54 UTC. Providers set Lium prices, so they move during the day; check the live listings before you rent.Connect Claude Code or Codex to Lium
Create a key at lium.io/api-keys, give it a daily budget and add credit.
In the terminal you use to launch Claude Code or Codex, set your API key:
bash
export LIUM_API_KEY="your_lium_api_key"Claude Code
bash
claude mcp add --transport http lium https://docs.lium.io/mcp --header "Authorization: Bearer $LIUM_API_KEY"Codex CLI (
~/.codex/config.toml)toml
[mcp_servers.lium]
url = "https://docs.lium.io/mcp"
bearer_token_env_var = "LIUM_API_KEY"Then ask the agent to find the cheapest free H100, rent it and remove it when the job is done. Step-by-step setups: Give Claude Code a GPU and Give Codex a GPU.
Choose your MCP server
If you want an agent to price a job before anyone signs up, and to stay inside a budget you set, Lium is the one built for it. Add the Lium MCP server to your agent and browse the live GPU listings.
Run it on Lium
Billed per second, only while your pod runs.
Get new posts by e-mail: Subscribe to the Lium newsletter
