Leeroopedia

Leeroopedia MCP

Give your AI coding agent access to curated ML/AI knowledge

[!TIP]

$20 free credit on sign-up. That's plenty of searches, plans, and diagnoses. Skip the guesswork on your next fine-tuning run or inference deployment. No credit card required. Get your API key →

What is Leeroopedia?

Your ML & AI Knowledge Wiki. Learnt by AI, built by AI, for AI.

Expert-level knowledge across the full ML & AI stack: fine-tuning and distributed training, inference serving and GPU kernel optimization, building agents and RAG pipelines. 1000+ frameworks and libraries, all in one place.

This MCP server turns your AI coding agent (Claude Code, Cursor, Claude Desktop, ChatGPT, OpenAI Codex, ...) into an ML/AI expert engineer.

Browse the full knowledge base at leeroopedia.com.

Want to go end-to-end?

Leeroopedia gives your agent the knowledge. Kapso gives it the ability to act on it: research, experiment, and deploy. Together: a complete ML/AI engineer agent.

Connect to Your Agents

Use our hosted server for zero-setup. Just paste this URL into any MCP client that supports remote servers:

https://mcp.leeroopedia.com/mcp?token=kpsk_your_key_here

Or see the per-client guides below for detailed instructions (including local setup).

[card] Claude Code

Set up with Claude Code

[card] Cursor

Set up with Cursor

[card] Claude Desktop

Set up with Claude Desktop

[card] OpenAI Codex

Set up with OpenAI Codex

[card] ChatGPT

Set up with ChatGPT

Benchmarks

We measured the effect of Leeroopedia MCP on real ML tasks:

• ML Inference Optimization. Write CUDA/Triton kernels for 10 KernelBench problems. 2.11x geomean speedup vs 1.80x (+17%), with/without Leeroopedia MCP.

• LLM Post-Training. End-to-end SFT + DPO + LoRA merge + vLLM serving + IFEval on 8×A100. 21.3 vs 18.5 IFEval strict-prompt accuracy, 34.6 vs 30.9 strict-instruction accuracy, 272.7 vs 231.6 throughput.

• Self-Evolving RAG. Build a RAG service that automatically improves itself over multiple rounds. 45.16 vs 40.51 Precision@5, 40.32 vs 35.29 Recall@5, in 52 vs 62 min wall time.

• Customer Support Agent. Multi-agent triage system classifying 200 tickets into 27 intents. 98 vs 83 benchmark performance, 11s vs 61s per query.

[card] See Full Benchmark Results

Detailed results, analysis, and replication instructions for all 4 benchmarks

Available Tools

The server provides 8 agentic tools: search, plan, review, verify, diagnose, hypothesize, query hyperparameters, and retrieve pages.

[card] Tools Overview

See all 8 tools with parameters and usage

Quick Links

[card] Claude Code

Connect in 2 minutes

[card] Cursor

Connect in 2 minutes

[card] Claude Desktop

Connect in 2 minutes

[card] OpenAI Codex

Connect in 2 minutes

[card] ChatGPT

Connect in 2 minutes

[card] Benchmarks

See the results

[card] Tools Overview

All 8 tools explained

This documentation is built and hosted on Mintlify, a developer documentation platform.