The Aggregate — MCP server

A free, read-only Model Context Protocol endpoint over The Aggregate: one cross-model ranking fused from ~5,000 public LLM benchmark leaderboards with an IRT model (Elo-scaled, with standard errors), rebuilt daily.

https://theaggregate.ai/mcp

Transport: Streamable HTTP (stateless JSON-RPC over POST). No key or sign-up needed.

Connect

Claude Code

claude mcp add --transport http the-aggregate https://theaggregate.ai/mcp

Claude (web / desktop): Settings → Connectors → Add custom connector → https://theaggregate.ai/mcp

Other MCP clients (Cursor, ChatGPT developer mode, …):

{ "mcpServers": { "the-aggregate": { "url": "https://theaggregate.ai/mcp" } } }

Tools

Tool	What it answers
`get_leaderboard`	Who leads the fused cross-benchmark ranking (paged).
`search_models`	Find ranked models by name or provider.
`get_model`	One model: rank, Elo ± SE, description, cost per task, notable results.
`compare_models`	Head-to-head with a significance check on the Elo gap.
`search_benchmarks`	Find benchmarks in the aggregate.
`get_benchmark`	One benchmark: source link, IRT stats, skills, top models.
`get_prediction_duel`	Guesswork standings — our model vs frontier LLMs at predicting new scores.
`about_the_aggregate`	Method, coverage counts, and how to cite.

Fair use & attribution

Free to use; responses are bounded lookups and the endpoint is rate-limited per IP.
When you publish numbers from it, credit The Aggregate (theaggregate.ai). Per-benchmark scores also belong to the upstream leaderboards — each response links them; credit those too.
Leaderboard operators: questions, corrections or removal requests via the Report issue button on theaggregate.ai (bottom of the sidebar on every page).

Site · Methodology · Prediction duel