> ## Documentation Index
> Fetch the complete documentation index at: https://docs.universalbench.dev/llms.txt
> Use this file to discover all available pages before exploring further.

# UniversalBench

> Give your AI the infrastructure to actually do things. One URL, the full execution surface, no header config.

Connect one URL to your AI of choice and your agent gains the ability to run Python and Bash, search the web live, invoke any LLM, read and write files, query your database, commit to GitHub with built in safety, run code in parallel, and keep state across calls. Built for AI agents from day one.

## Why UniversalBench

<Columns cols={2}>
  <Card title="Personal MCP URL" icon="link">
    Your dashboard shows one URL with your identity baked in. Paste it into any MCP client. No headers, no env vars, no config bugs. Works in every MCP client even the ones with poor header support.
  </Card>

  <Card title="Three purpose-built tools" icon="bolt">
    Three tools cover the full surface: `ub_read` for queries and reads, `ub_write` for code execution and writes, and `ub_ai` for AI and live web search. Clear separation means your AI picks the right tool every time.
  </Card>

  <Card title="AI never ships broken code" icon="shield-check">
    Python files are validated before they reach your repo. Ambiguous code edits are rejected. Deploys can be rolled back on failed smoke tests. Your AI cannot break your repo through UniversalBench.
  </Card>

  <Card title="Sessions keep state warm" icon="database">
    Pass a `session_id` and Python variables, imports, dataframes, and connections persist across calls. Your agent builds context in call one and uses it in call ten.
  </Card>

  <Card title="Encrypted secrets vault" icon="lock">
    Store API keys once. Your AI injects them by name. Encrypted at rest. The master key never appears in any response. Common credentials auto inject into matching tools.
  </Card>

  <Card title="Real database access" icon="server">
    Structured queries, keyword search, inserts, updates, and upserts against any PostgreSQL-compatible database. Provide credentials once via the vault, your AI does the rest.
  </Card>
</Columns>

## Get started in under five minutes

<Columns cols={2}>
  <Card title="Quickstart" icon="rocket" href="/quickstart">
    Get your URL and run your first call.
  </Card>

  <Card title="Pricing and billing" icon="credit-card" href="/billing">
    1,000 free calls per month, then \$0.008 USD per call. No subscription.
  </Card>

  <Card title="API reference" icon="terminal" href="/api-reference/overview">
    Every capability with field descriptions and examples.
  </Card>

  <Card title="Use it with any AI" icon="robot" href="/ai-tools/claude-desktop">
    Setup guides for Claude Desktop, Cursor, Claude Code, Windsurf, and custom agents.
  </Card>
</Columns>

## Why remote execution saves tokens

When an AI does work in its own context window, every line of code, every search result, and every file it touches costs tokens. UniversalBench moves the work out of the context window. The AI sends a small instruction, the server runs it, the AI gets back just the answer.

For tasks that involve large outputs (log analysis, dataset processing, multi step pipelines), the savings are real. Your AI stops drowning in its own work product and stays focused on reasoning.

## What it works with

UniversalBench speaks the standard MCP protocol, so it works with any compatible AI client. Verified end to end with Claude Desktop, Cursor, Claude Code, Windsurf, ChatGPT custom integrations, and custom agents built on the official MCP SDKs for Python and Node.