Mahal Systems

Overview › Benchmark catalog

MCP-Universe

Fits agents that drive real third-party services (maps, GitHub, browser, market data) over long tool-calling sessions. Borrow its evaluator pattern (extract a value from the output, compare with an operator) and its "dynamic evaluators" that fetch live ground truth for time-sensitive tasks instead of freezing an answer.

What it asks the agent to do

  • find places and routes with a maps server
  • inspect or change a GitHub repository
  • answer a financial analysis question from market data
  • build a 3D object in Blender through its MCP server
  • complete a browser automation task
  • answer a time-sensitive question with web search

What to borrow from it

grading designcase templatebaseline

Caveats

Needs paid API keys (SerpAPI, Google Maps, GitHub, Notion) and a local Blender install; GitHub tasks perform real writes so a throwaway account is required; 231 hand-written tasks across six domains only; 57% of tasks depend on live data that changes.

Details

Workloadstool calling, web browsing, finance, devops, research
Interactionautonomous long horizon
Environmentmixed
End state checkedNo
Tools11 real MCP servers (Google Maps, GitHub, financial data, Blender, Playwright, web search, Notion)
Gradingprogrammatic check, exact match
Size231
Datamixed
Personal datapossible
LicenseApache-2.0
Harnessown repo
Statusactive

Links