Cerebrium agent skills

Official Agent Skills and hosted docs MCP for

Cerebrium, the serverless GPU and CPU platform for real-time AI

workloads. The skills teach a coding agent to write a valid `cerebrium.toml`, pick hardware and a

region, deploy, call the endpoint, and debug the result without leaving the terminal. The MCP

gives it live search over the documentation.

Works with Claude Code, Codex, Cursor, GitHub Copilot, Windsurf, Cline and anything else that

reads the Agent Skills format.

Install

Any agent (installs the skills into every detected harness):

npx skills add CerebriumAI/cerebrium-skills -g -y
npx add-mcp https://cerebrium.ai/docs/mcp -n cerebrium-docs -g -y

Claude Code (one plugin bundles the skills and the docs MCP):

/plugin marketplace add CerebriumAI/cerebrium-skills
/plugin install cerebrium@cerebrium

Codex:

codex plugin marketplace add CerebriumAI/cerebrium-skills

Gemini CLI (one extension bundles the skills and the docs MCP):

gemini extensions install https://github.com/CerebriumAI/cerebrium-skills

Cursor loads the repository directly as an Agent Plugin: the root

`plugin.json` and `mcp.json` are the open-standard manifests, so no Cursor-specific install step

is needed.

Or agent-driven: paste this into the agent of your choice.

Install the Cerebrium agent toolkit following instructions from
github.com/CerebriumAI/cerebrium-skills: use `npx skills add` and `npx add-mcp`,
global and auto-confirmed for all agents (-g -y).

Then create an account and authenticate the CLI once:

# create an account at https://dashboard.cerebrium.ai (also where API keys are created)
pip install cerebrium   # or: brew tap cerebriumai/tap && brew install cerebrium
cerebrium login

Compute is billed per second: see pricing for current rates

and any starting credit.

In CI, skip `login` and set `CEREBRIUM_SERVICE_ACCOUNT_TOKEN` instead.

What is in here

One skill, structured for progressive disclosure: agents load the ~120-line core

(`skills/cerebrium/SKILL.md`, the workflow, the safety rules, the endpoint shapes) when a

Cerebrium task appears, and pull in a reference file only when the task needs it.

The docs MCP (hosted at `https://cerebrium.ai/docs/mcp`) is read-only: documentation search and

filesystem, no account access, no key needed. It is declared twice on purpose, because the two

formats are not interchangeable: `.mcp.json` is the Claude Code convention, and `mcp.json` at the

repository root is the path the Agent Plugins standard requires, with that standard's

`streamable-http` transport name.

The plugin metadata is likewise declared per host: `.claude-plugin/` for Claude Code,

`.agents/plugins/` for Codex, root `plugin.json` for Agent Plugins clients such as Cursor, and

`gemini-extension.json` for the Gemini CLI. `server.json` is the MCP registry entry.

Try it

• "Deploy this FastAPI app to Cerebrium on an H100 with auth enabled, then follow the logs."
• "My Cerebrium app queues requests under load. Work out why and fix the config."
• "Which GPU and region fit a 13B vLLM model on Cerebrium, and what should the cerebrium.toml be?"

Related surfaces

• Any docs page as markdown: append `.md` to the URL
• Page index: `https://cerebrium.ai/docs/llms.txt`
• Runnable examples: CerebriumAI/examples

Accuracy

Every default, enum, flag and signature here was read out of the CLI source and the API

validator rather than copied from a docs page. See CONTRIBUTING.md for where

to check each kind of claim, and keep it that way.

MIT licensed.