Quickstart - UniversalBench

There are two ways to connect UniversalBench. Choose the one that matches how you are using it.

Option A: Connect via the Anthropic Connectors Directory (recommended for Claude.ai)

If you found UniversalBench in the Claude.ai connectors directory, just click Connect. The OAuth flow handles everything. You will be asked to sign in or create an account, then Claude.ai is connected automatically. No URL to copy, no config file to edit. That is it. Skip to Run your first call.

Option B: Connect with a personal URL (Claude Desktop, Cursor, Claude Code, custom agents)

For any MCP client outside of Claude.ai, sign up at universalbench.dev. After you verify your email, your dashboard shows your personal MCP URL:

https://universalbench-mcp.penantiaglobal.workers.dev/u/ubk_xxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxx

Treat this URL like a password. Anyone with it can make calls billed to your account.

Claude Desktop
Cursor
Claude Code
Direct HTTP

Edit ~/Library/Application Support/Claude/claude_desktop_config.json on macOS or %APPDATA%\Claude\claude_desktop_config.json on Windows:

{
  "mcpServers": {
    "universalbench": {
      "url": "https://universalbench-mcp.penantiaglobal.workers.dev/u/ubk_xxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxx"
    }
  }
}

Restart Claude Desktop. UniversalBench appears in the tool list within seconds.

Add to your Cursor MCP config:

{
  "mcpServers": {
    "universalbench": {
      "url": "https://universalbench-mcp.penantiaglobal.workers.dev/u/ubk_xxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxx"
    }
  }
}

From any directory:

claude mcp add universalbench https://universalbench-mcp.penantiaglobal.workers.dev/u/ubk_xxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxx

Call the API directly. The URL is the auth, no header needed:

curl -X POST https://universalbench-mcp.penantiaglobal.workers.dev/u/ubk_your_key \
  -H "Content-Type: application/json" \
  -H "Accept: application/json, text/event-stream" \
  -d '{"jsonrpc":"2.0","id":1,"method":"tools/call","params":{"name":"ub_write","arguments":{"code":"print(sum(range(100)))"}}}'

Run your first call

Ask your AI something that requires execution:

Use UniversalBench to find all prime numbers under 10,000 and tell me the largest gap between consecutive primes.

Your AI routes the heavy lifting to UniversalBench, which runs the Python and returns just the answer. Your AI synthesizes a clean response from a small result instead of generating thousands of tokens in its own context.

What to try next

Web search with live data

Ask your AI to search the web for current information. UniversalBench returns clean snippets with sources.

Invoke any LLM

Use invoke_llm to call models from any provider on demand. Compare outputs, route by cost, or run a cheap model as a first pass.

Keep state across calls

Pass session_id so your Python variables and imports persist between calls. Build a dataframe once, query it many times.

Store secrets safely

Use secrets_vault to store API keys encrypted server side. Your AI uses them by name without ever seeing the value.

Pricing

You start with 1,000 free calls every month, no credit card required. After that, top up your wallet starting at $5 and pay $0.008 per call. No subscription, funds never expire. See the billing page for the full breakdown.

Need help

Email hello@universalbench.dev. The dashboard at universalbench.dev shows your usage, wallet balance, and lets you rotate your URL if needed.

​Option A: Connect via the Anthropic Connectors Directory (recommended for Claude.ai)

​Option B: Connect with a personal URL (Claude Desktop, Cursor, Claude Code, custom agents)

​Run your first call

​What to try next