@autoinference/skill

MCP server skill for autoinference — peak inference performance, zero ML engineering.

Drop in any frontier LLM (DeepSeek V4 Pro, Kimi 2.6, GLM 5.1, ...) on any hardware (8× B300, 256× H200 multi-node, AMD MI355X, ...) and the autoinference agentic pipeline picks the optimal inference engine, tunes every config, synthesises custom kernels, and stress-tests the result.

This package exposes that pipeline as an MCP server so any MCP-aware coding agent — Claude Code, Codex, Cursor, Antigravity — can call into it directly from a chat.

⚠️ v0.0.1-alpha · placeholder release. The MCP surface installs and connects, but the underlying agent pipeline ships in a later release. Today's tool calls return friendly placeholder responses describing what the call would do.

Install

Universal (any MCP client)

npx skill add @autoinference/skill

Or with the canonical npm npx:

npx -y @autoinference/skill

Per-editor

| Editor | Command | | ------------ | ------------------------------------------------------------- | | Claude Code | claude mcp add autoinference | | Codex | codex mcp add autoinference | | Cursor | cursor mcp add autoinference | | Antigravity | antigravity mcp install autoinference |

For Cursor's ~/.cursor/mcp.json:

{
  "mcpServers": {
    "autoinference": {
      "command": "npx",
      "args": ["-y", "@autoinference/skill"]
    }
  }
}

Tools exposed

| Tool | Description | | -------------------------- | --------------------------------------------------------------------------------- | | autoinference.deploy | Deploy an LLM optimally on user-owned hardware. Returns a placeholder in alpha. | | autoinference.benchmark | Run a load/latency/cost benchmark against a deployment. Placeholder in alpha. | | autoinference.summary_url| Return the wandb-style dashboard URL for a deployment. Placeholder in alpha. |

Status

• npm package: @autoinference/skill
• Source: github.com/autoinference/autoinference
• Website: autoinference.org

License

MIT