Use with Cursor

Cursor speaks MCP natively. Adding UniversalBench takes about 30 seconds because there is no header config to fight with. The connection is one URL.

Setup

Get your personal MCP URL

Sign up at universalbench.dev and copy your URL from the dashboard. It looks like https://universalbench-mcp.penantiaglobal.workers.dev/u/ubk_...

Open Cursor MCP settings

Open Settings with Ctrl+, (Windows or Linux) or Cmd+, (macOS), then go to Tools & MCP in the sidebar and click Add Custom MCP. Cursor opens your mcp.json file (at ~/.cursor/mcp.json).

Paste the config

Replace the contents of mcp.json with this, using your own URL:

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

Notice there is no headers block. The URL itself is the auth. Save the file.

Confirm the connection

Back in Tools & MCP, universalbench appears under User MCP Servers with a green dot, and three tools are listed: ub_read, ub_write, and ub_ai. The connection is usually instant. If it does not appear, reopen the settings panel or restart Cursor.

The three tools

UniversalBench exposes three tools, grouped by what they do:

Tool	What it does
`ub_read`	Read-only work: read files, run read-only SQL (`db_select`, `db_query`, `db_search`), read from GitHub, check account status
`ub_write`	Anything that changes state: run code and bash, write files, install packages, run parallel jobs, push to GitHub with validation, safe deploys, write to your database, store secrets in the vault, make outbound HTTP calls
`ub_ai`	Web search and LLM calls, with cost caps on by default

What you can do now

Ask the Cursor Composer things that need execution:

Run a quick analysis on the test runs in /tmp/results.csv. Group failures by file and tell me the top three offenders.

Cursor uses ub_write with its code capability to read the CSV with pandas, group the failures, and return the analysis as plain text. No CSV dump in the context window. Or:

Use the cheapest available LLM via UniversalBench to draft a commit message for my staged diff.

Cursor uses ub_ai with invoke_llm to route to a cheap model, saving tokens compared to using the editor’s primary model for low value text generation.

When to reach for UniversalBench in Cursor

UniversalBench is not a replacement for Cursor’s local tools. For editing files in your repo and running local tests, Cursor’s own tools are faster. UniversalBench earns its place when you want:

The agent to query or change your database or GitHub using credentials stored once in the vault, without pasting tokens into chat
Validated pushes and safe deploys with rollback, rather than commit and hope
An isolated cloud runtime when you do not want agent code running on your machine, or your local shell is restricted
Cost-capped LLM and web calls for small bounded subtasks

For plain local edits, stick with Cursor’s built in tools.

Tips

For multi step pipelines that need state, pass a session_id so imports and variables persist across calls.

Troubleshooting

Problem	Fix
Tools do not appear after saving	Reopen Tools & MCP, or restart Cursor. The server should show “universalbench” with a green dot
401 or 403 errors	The token after `/u/` has a typo. Paste the URL again from the dashboard
Slow first call	The first call after idle has a small cold start. Subsequent calls are fast

Get started

Use it with any AI

Patterns for AI agents

Advanced

Setup

The three tools

What you can do now

When to reach for UniversalBench in Cursor

Tips

Troubleshooting

​Setup

​The three tools

​What you can do now

​When to reach for UniversalBench in Cursor

​Tips

​Troubleshooting

Setup

The three tools

What you can do now

When to reach for UniversalBench in Cursor

Tips

Troubleshooting