MCP Atlas
Can the model find, combine and drive the right tools through the Model Context Protocol to solve a realistic task?
What does it measure?
MCP Atlas measures how well a language model discovers tools, fills in parameters correctly, coordinates multiple servers and grounds its answers in tool output. The benchmark uses 1,000 human-written tasks across 36 real MCP servers and 220 tools. Each task also includes distractor tools that look relevant but are not needed.
How to read the score
Claims-based scoring with partial credit. Each task defines factual claims the answer must contain. The model earns points for each correct claim, plus diagnostics on tool discovery, parameterization, syntax, error recovery and efficiency.
Example task
Use the Spotify MCP server to fetch an artist's top 5 tracks and the GitHub MCP server to create an issue with that data as its content.
What to watch out for
MCP Atlas tests single-turn tasks only: the model gets one instruction and has to pick and call the right tools in one go. Multi-turn conversations are not tested. The 500 tasks in the public leaderboard are a subset of the full 1,000.