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.

Claude Code is Anthropic’s terminal native coding agent. It supports MCP servers out of the box, so adding UniversalBench is one command.

​
Setup

1

Get your personal MCP URL

Sign up at universalbench.dev and copy your URL from the dashboard.
2

Add UniversalBench

From any directory:
claude mcp add universalbench https://mcp.universalbench.dev/u/ub_live_xxxxxxxxxxxxxxxxxxxxx
No --header flag needed. The URL carries your auth.Or edit ~/.claude/mcp_config.json directly:
{
  "mcpServers": {
    "universalbench": {
      "url": "https://mcp.universalbench.dev/u/ub_live_xxxxxxxxxxxxxxxxxxxxx"
    }
  }
}
3

Verify it connected

claude mcp list
UniversalBench should appear with status “connected”.

​
What you can do now

Ask Claude Code things that need real execution:
Find every TODO in the codebase, group by file owner, and write a summary to TODO_OWNERSHIP.md. Push it on a branch called chore/todo-audit.
Claude Code uses UniversalBench to grep across your repo at remote speed, use Python to group and format the results, then call validate_and_push to ship the new file on a branch with syntax gates. Or for pipelines:
Check the last 24 hours of logs for any 500 errors. If there are more than 5, page on call.
Claude Code chains db_select (read logs), parallel_blocks (analyze in parallel), and proxy_http (send the PagerDuty alert).

​
Tips

Add --allow-tools universalbench__* to your claude invocation to skip approval prompts for UB calls when you trust the workflow.
Claude Code’s session memory is local. UniversalBench’s session_id adds remote session state so long running pipelines survive across CLI invocations.

​
Troubleshooting

ProblemFix
claude mcp list shows “disconnected”Check claude mcp logs universalbench for the underlying error
401 or 403 errorsURL has a typo. The token after /u/ must match exactly what your dashboard shows
Tool calls timeoutLong running calls should use session_id plus async patterns rather than blocking