Fine-tune LLMs from any AI assistant with the BiOS MCP server

The BiOS MCP server gives your AI assistant 38 tools that cover the whole fine-tuning workflow: upload datasets, pick a model, launch and monitor training, and deploy the result for inference. It is built on the Model Context Protocol, an open standard, so one npm package works across every MCP client.

How it works

Desktop clients launch the server as a small local process and talk to it over MCP. The server calls the BiOS API with your key. No proxy sits in between.

Your AI assistant

Claude Code, Cursor, Windsurf, Cline, VS Code, Zed, Devin

MCPstdio

BiOS MCP server

npx @usfbios/mcp-server
runs on your machine

HTTPSAPI key

BiOS platform

Datasets, training jobs, checkpoints, inference

The server runs locally, so your API key stays on your machine and never passes through the model provider.

Hosted connector

Web and mobile AI apps (claude.ai, ChatGPT) skip the local server and connect at https://mcp.usbios.ai/mcp. You sign in with your BiOS account and approve the connection, so there is no install and no API key. Desktop tools can use either the hosted address or the local npx server below.

Setup guides by client

Every guide takes about two minutes: paste one config block, restart, done. Pick your client:

Claude Code

Terminal coding agent by Anthropic

Claude Desktop

Claude app for macOS and Windows

Claude on the web

claude.ai in the browser and the mobile apps

ChatGPT

ChatGPT and the OpenAI Codex CLI

Cursor

AI code editor

Windsurf

AI code editor with Cascade agent

Cline

Open-source VS Code agent extension

VS Code

GitHub Copilot agent mode

Zed

High-performance editor with Agent Panel

Devin

Autonomous software engineer by Cognition

Gemini CLI

Google terminal coding agent

Quick start (any client)

1. Create an API key

In your dashboard, open API Keys, create one, and copy the sk_live_ value.

2. Add the server

Paste the config below into your client. It runs the server with npx, so there is nothing to install first.

3. Ask in plain English

The assistant discovers the tools on its own and asks for approval before each call.

mcp config (most clients)
{
  "mcpServers": {
    "usfbios": {
      "command": "npx",
      "args": ["-y", "@usfbios/mcp-server"],
      "env": {
        "USFBIOS_API_KEY": "sk_live_your_api_key"
      }
    }
  }
}

Node.js 18 or newer is the only requirement. Exact file locations differ per client; the per-client guides above have them.

What the 38 tools cover

The full tool reference with every parameter lives in the API documentation. At a glance:

5

Datasets

Upload local files, import from HuggingFace, preview and manage

12

Training

Launch, monitor, stop, and resume jobs; metrics, logs, evals, checkpoints

11

Inference

Deploy fine-tunes, manage endpoints, chat with a deployed model

7

Models & GPU

Search 250+ base models, get configs and GPU recommendations with cost estimates

3

Account

Wallet balance, key introspection, and a built-in platform guide

Things you can say

“Search for Llama models under 13B parameters and recommend a GPU for LoRA fine-tuning”

“Upload ~/data/support-qa.jsonl and start an SFT job on Llama 3.1 8B with LoRA”

“What is the status of my running training jobs? Show the loss curve for the latest one”

“Deploy my finished fine-tune and give me the endpoint URL”

Frequently asked questions

What is the Model Context Protocol (MCP)?

MCP is an open standard that lets AI assistants call external tools. A client like Claude or Cursor connects to an MCP server, discovers its tools, and calls them during a conversation. The BiOS MCP server exposes the whole fine-tuning workflow as tools.

Which AI assistants work with the BiOS MCP server?

Any client that can run local MCP servers: Claude Code, Claude Desktop, Cursor, Windsurf, Cline, VS Code Copilot agent mode, Zed, Devin, the OpenAI Codex CLI, the Gemini CLI, and more. Web and mobile apps such as claude.ai and ChatGPT connect through the hosted connector at https://mcp.usbios.ai/mcp with a sign-in, no install needed.

Is the MCP server free?

Yes. The @usfbios/mcp-server package is free to install and run. You pay only for what you use on the platform itself, such as GPU time for training jobs, billed per second to your wallet.

Where does my API key go?

The server runs as a local process on your machine and sends the key only to the BiOS API over HTTPS. It never passes through the AI model provider. Keys are scoped, so you can restrict one to read-only access if you prefer.

Can the assistant spend money without asking me?

Every MCP client asks for your approval before running a tool by default. Cost-bearing tools like create_training_job also run preflight checks, and your wallet balance caps total spend.

Ready to try it?

Create a free account, grab an API key, and your assistant can launch its first fine-tuning job a few minutes from now.

Create accountFull tool reference