UniversalBench: run code, web, databases, and any LLM from your AI, safely
UniversalBench
AI has the intelligence. UniversalBench gives it the infrastructure to act.
Connect one URL to Claude, ChatGPT, Gemini, or any MCP-compatible AI, and behind it your AI can run code, search the web, read and write databases, call any LLM, and commit to GitHub, all before results reach the model. That filtering is what cuts up to 96.5 percent of tokens versus doing the work in chat.
Three guarantees, all live in code:
-
AI never ships broken code. Every code push is validated and confirmed to load before commit.
-
AI never burns your wallet. Every LLM call is cost-estimated before it runs, and calls over your ceiling are rejected.
-
AI cannot reach your internal network. Every request is checked against an internal-network blocklist.
What your AI can do through three tools (ub_read, ub_write, ub_ai):
-
Run code and shell commands in an isolated sandbox
-
Install packages and run tasks in parallel
-
Search the web and make safe HTTP calls
-
Read, write, and query databases with your own credentials
-
Call any major LLM with cost caps on by default
-
Read from and commit to GitHub, with every push validated first
-
Take screenshots and handle file delivery and conversion
-
Persist memory and state across sessions
-
Store credentials in an encrypted vault that unlocks the right tools automatically
Connect: https://universalbench-mcp.penantiaglobal.workers.dev/u/{api_key}
Free tier: 1,000 calls per month, no card required, then pay as you go.
Server Config
{
"mcpServers": {
"universalbench": {
"url": "https://universalbench-mcp.penantiaglobal.workers.dev/u/{api_key}"
}
}
}