# Leeroopedia MCP Server

Give your AI coding agent access to best-practices of ML and AI.

Leeroopedia MCP overview

--- > **$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 →](https://app.leeroopedia.com) ## 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, from fine-tuning and distributed training, to inference serving and GPU kernel optimization, to building agents and RAG pipelines. **1000+ frameworks and libraries**, all in one place. This MCP server turns your AI coding agent (Claude Code, Cursor, OpenAI Codex, ...) into an ML/AI expert engineer. Browse the full knowledge base at [leeroopedia.com](https://leeroopedia.com). ### Want to go end-to-end? Leeroopedia gives your agent the **knowledge**. [**Kapso**](https://github.com/leeroo-ai/kapso) gives it the **ability to act on it** : research, experiment, and deploy. Together: a complete ML/AI engineer agent. ## Benchmarks We measured the effect of Leeroopedia MCP on real ML tasks built by Claude Code. - **ML Inference Optimization**: Write CUDA/Triton kernels for 10 KernelBench problems. **2.11x** geomean speedup vs 1.80x (**+17%**), with/without Leeroopedia MCP. [→ results](examples/ml_inference_optimization/) - **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. [→ results](examples/llm_post_training/)

- **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. [→ results](examples/self_evolve_rag/)

- **Customer Support Agent**: Multi-agent triage system classifying 200 tickets into 27 intents. **98 vs 83** benchmark performance, **11s vs 61s** per query. [→ results](examples/customer_support_agent/)

## Quick Start ### 1. Get Your API Key 1. Go to [app.leeroopedia.com](https://app.leeroopedia.com) 2. Create an account or log in 3. Navigate to **Dashboard > API Keys** 4. Copy your API key (format: `kpsk_...`) ### 2. Use the hosted server (no installation needed) Just paste this URL into any MCP client that supports remote servers: ``` https://mcp.leeroopedia.com/mcp?token=kpsk_your_key_here ``` See the **[connect guides](https://docs.leeroopedia.com)** for all IDEs, running locally, and troubleshooting. ### 3. Configure Claude Code **Remote (no install):** ```bash claude mcp add --transport http leeroopedia "https://mcp.leeroopedia.com/mcp?token=kpsk_your_key_here" ``` **Local (via uvx):** Add to your `~/.claude.json` or project `.mcp.json`: ```json { "mcpServers": { "leeroopedia": { "command": "uvx", "args": ["leeroopedia-mcp"], "env": { "LEEROOPEDIA_API_KEY": "kpsk_your_key_here" } } } } ``` > **Getting `spawn uvx ENOENT`?** Your IDE can't find `uvx` in its PATH. Run `which uvx` (or `where uvx` on Windows) in your terminal to get the full path, then use it in your config: > ```json > "command": "/home/username/.local/bin/uvx" > ``` > Common locations: `~/.local/bin/uvx` (Linux), `~/.local/bin/uvx` (macOS curl install), `/opt/homebrew/bin/uvx` (macOS Homebrew). ### 4. Configure Cursor **Remote (no install):** Add to your Cursor settings (`.cursor/mcp.json`): ```json { "mcpServers": { "leeroopedia": { "url": "https://mcp.leeroopedia.com/mcp?token=kpsk_your_key_here" } } } ``` **Local (via uvx):** ```json { "mcpServers": { "leeroopedia": { "command": "uvx", "args": ["leeroopedia-mcp"], "env": { "LEEROOPEDIA_API_KEY": "kpsk_your_key_here" } } } } ``` > **Getting `spawn uvx ENOENT`?** See the tip in [Configure Claude Code](#3-configure-claude-code) above. ### 5. Configure OpenAI Codex Run the CLI command: ```bash codex mcp add leeroopedia --env LEEROOPEDIA_API_KEY=kpsk_your_key_here -- uvx leeroopedia-mcp ``` Or add to your `~/.codex/config.toml`: ```toml [mcp_servers.leeroopedia] command = "uvx" args = ["leeroopedia-mcp"] [mcp_servers.leeroopedia.env] LEEROOPEDIA_API_KEY = "kpsk_your_key_here" ``` > **Getting `spawn uvx ENOENT`?** See the tip in [Configure Claude Code](#3-configure-claude-code) above. ### 6. Optional: Add the Agent Skill File For coding agents that support project-level instruction files (Claude Code, Cursor, Codex), you can add an optional skill file that teaches the agent **when and how** to pick the right Leeroopedia tool for each situation. This improves tool selection, encourages parallel searches, and provides canonical workflows for common tasks like "How do I implement X?" or "My run crashed". Download [`SKILL.md`](https://raw.githubusercontent.com/Leeroo-AI/leeroopedia-mcp/main/SKILL.md) and place it in the appropriate location for your agent: | Agent | Where to place the file | Format | |-------|------------------------|--------| | **Claude Code** | Project root or `CLAUDE.md` (paste the content) | Markdown | | **Cursor** | `.cursor/rules/leeroopedia.md` (or add as an Agent Skill) | Markdown | | **Codex** | Project root `AGENTS.md` (paste the content) | Markdown | > This step is optional. The MCP tools work without the skill file. The skill file just helps the agent make better tool choices automatically. ## Documentation > **[docs.leeroopedia.com](https://docs.leeroopedia.com)** — full setup guides, environment variables, troubleshooting, connect guides for every IDE, and detailed tool usage. ## Contributing We welcome contributions! Please see our [Contributing Guide](CONTRIBUTING.md) for details on how to get started. This project follows our [Code of Conduct](CODE_OF_CONDUCT.md). ## License This project is licensed under the [MIT License](LICENSE).