MCP Server & SDK

BiOS provides three ways to automate fine-tuning programmatically: an MCP server for AI assistants (Claude, Cursor, Windsurf), a TypeScript SDK and Python SDK for application code, and a REST API (documented in previous sections). All three use API keys created from your dashboard.

API Keys

All programmatic access requires an API key. Each key is scoped to a workspace with fine-grained permissions. You only need the API key itself. Org and workspace are resolved automatically.

1. Go to your dashboard and click API Keys in the sidebar
2. Click Create API Key, name it, and choose a permission preset (Read Only, Training, Full Access, or Custom)
3. Optionally set an expiration (30, 60, 90, 180 days, or 1 year)
4. Copy the key immediately; it is only shown once
5. Set the key as USFBIOS_API_KEY. Nothing else is needed

Permission Scopes

Call GET /api/api-keys/introspect with your key to see exactly what it can do: permissions, workspace, allowed tools, and rate limits.

MCP Server

The Model Context Protocol (MCP) is an open standard that lets AI assistants call external tools. The BiOS MCP server gives your AI assistant 38 tools covering the entire fine-tuning workflow, from datasets to training to inference. It works with Claude Desktop, Claude Code, Cursor, Windsurf, Cline, VS Code, Zed, Devin, and any MCP-compatible client.

Web and mobile apps such as claude.ai and ChatGPT connect through the hosted connector at https://mcp.usbios.ai/mcp instead: you sign in with your BiOS account and approve the connection, with no install and no API key.

For local setups, install globally or run with npx:

npm install -g @usfbios/mcp-server
# or run directly:
npx @usfbios/mcp-server

Client Setup Guides

We publish a dedicated step-by-step guide for every major MCP client, each with the exact config file location, verification steps, and troubleshooting:

Or start from the MCP server overview for the quick start that applies to any client.

Configuration

Add to your AI client's configuration.

Claude Desktop / Claude Code

Add this to your claude_desktop_config.json or the project's .mcp.json:

{
  "mcpServers": {
    "usfbios": {
      "command": "npx",
      "args": ["@usfbios/mcp-server"],
      "env": {
        "USFBIOS_API_KEY": "sk_live_your_api_key"
      }
    }
  }
}

Cursor / Windsurf

Add the same configuration to your .cursor/mcp.json or the IDE's MCP settings panel.

Environment Variables

Available Tools

The MCP server exposes 38 tools organized by category. The most used ones are documented below; the inference tools mirror the Inference API. Your AI assistant can call any of them:

Usage Examples

Once connected, you can ask your AI assistant to perform fine-tuning tasks in natural language. Here are some example prompts:

End-to-End Fine-Tuning Example

This example shows the complete flow, from finding a model to launching a training job, as a single conversation with your AI assistant:

User: I have a customer support dataset at ~/data/support-qa.jsonl.
      Fine-tune Llama 3.1 8B on it using SFT with LoRA. Use default hyperparameters.

Assistant calls: search_models(query="Llama 3.1 8B")
  → Found: meta-llama/Llama-3.1-8B-Instruct (8B params)

Assistant calls: upload_dataset(file_path="~/data/support-qa.jsonl")
  → Uploaded: ds_abc123 (12,450 rows, JSONL, validated)

Assistant calls: get_gpu_pricing()
  → Recommended: A100 80GB at $2.49/hr for 8B model

Assistant calls: create_training_job(
    model="meta-llama/Llama-3.1-8B-Instruct",
    dataset_id="ds_abc123",
    method="sft",
    adapter="lora",
    gpu_type="A100_80GB"
  )
  → Job launched: job_def456 (estimated 2-3 hours, ~$6.50)

TypeScript / JavaScript SDK

For application code, use the @usfbios/sdk package:

npm install @usfbios/sdk

import { USFBios } from '@usfbios/sdk';

const client = new USFBios({
  apiKey: 'sk_live_...',        // from Dashboard → API Keys (that's all you need)
});

// Discover what your key can do
const info = await client.introspect();
console.log(info.scopes);       // ['models:read', 'datasets:write', ...]
console.log(info.org.name);     // "Arpit's Org"

// Search models
const models = await client.models.search({ query: 'llama 8B' });

// Upload a dataset
const ds = await client.datasets.upload({
  filePath: './support-qa.jsonl',
  name: 'Customer Support QA'
});

// Check wallet balance
const balance = await client.wallet.getBalance();

// Get GPU recommendation
const gpu = await client.gpu.getRecommended(
  'meta-llama/Llama-3.1-8B-Instruct'
);

// Launch fine-tuning
const job = await client.training.create({
  model: 'meta-llama/Llama-3.1-8B-Instruct',
  datasetId: ds.id,
  method: 'sft',
  adapter: 'lora',
});

// Monitor progress
const status = await client.training.get(job.id);
const metrics = await client.training.getMetrics(job.id);

// Stop / resume
await client.training.stop(job.id);
await client.training.resume(job.id);

The SDK supports both API key and JWT authentication. It works in Node.js, edge runtimes, and any JavaScript environment with native fetch.

All SDK Methods

Python SDK

For Python applications, use the usfbios package:

pip install usfbios

from usfbios import USFBios

client = USFBios(api_key="sk_live_...")

# Search models
models = client.models.search(query="llama 8B")

# Upload a dataset
ds = client.datasets.upload(file_path="./data.jsonl", name="My Dataset")

# Launch fine-tuning
job = client.training.create(
    model="meta-llama/Llama-3.1-8B-Instruct",
    dataset_id=ds.id,
    method="sft",
    adapter="lora",
)

# Monitor progress
metrics = client.training.get_metrics(job.id)
print(f"Loss: {metrics.current_loss}")

The Python SDK mirrors the TypeScript SDK method-for-method using snake_case naming. Async support is available via AsyncUSFBios. See the full Python SDK docs in the sidebar for complete usage, type hints, and async examples.