Skip to main content

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.

Your AI agent sees UniversalBench as a set of tools with names, descriptions, input schemas, and examples. This is what the AI uses to decide whether and how to call each capability.

​
What is in the AI’s context

When your AI client (Cursor, Claude Desktop, your custom agent) connects to UniversalBench, it pulls a manifest like this: 
{
  "tools": [
    {
      "name": "workbench_execute",
      "description": "Execute Python or Bash code in a sandboxed environment. Returns stdout, stderr, and execution time.",
      "inputSchema": { "/* 33 fields with descriptions */": "" }
    }
  ]
}
Each of UniversalBench’s 33 capabilities ships with a clear name (db_select, safe_deploy, secrets_vault), a one line description of what it does, input field descriptions for every parameter, and an example object showing typical usage.

​
Why this matters

Your AI’s decision quality depends on this metadata. UniversalBench invests heavily in three areas: Descriptions that pass the stranger test. Each tool description has to make sense to an AI that has never used UB before. We do not assume context. Examples that compile. Every example object is real, copy pasteable JSON. No placeholders that break when used literally. Schemas that prevent errors. Required fields are marked required. Enums constrain values to what actually works. Pattern validators catch malformed inputs before they hit the workbench.

​
How to help your AI choose well

​
1. Name the capability you want

If you want a specific tool, name it. AIs are good at fuzzy matching but explicit is always faster.
Use db_select with a filter on created_at > '2026-01-01'
instead of
Get me the recent records

​
2. Pass context once

If your task involves the same database or repo across many calls, mention it at the top of the conversation. Most MCP clients carry context across turns.
We are working in the cust_prod database and the nikhilgogulwar/app repo for this session.

​
3. Trust the rollback

safe_deploy and validate_and_push have built in safety. You do not need to add “be careful” to your prompt. Be specific about what you want and let UB handle the safety.

​
What your AI does not see

  • Your personal MCP URL token (the client holds the URL, never echoes the token back into the AI’s context)
  • Other customers’ data (every call is scoped to the customer your URL belongs to)
  • The workbench’s internal state from previous calls (unless you pass session_id)
  • Secret values stored in secrets_vault (only their names appear in tool calls)
This is by design. Your AI agent gets enough context to be useful, never enough to leak. Because auth lives in the URL path and not in your prompts, even a fully recorded conversation transcript does not expose your credentials.