MCP Server

Use with Claude Code

LLM Dojo exposes a knowledge base MCP server. Connect it to Claude Code and any agent can query the full curriculum — finding the right notebook, concept, or Colab link for any LLM question.

01 — Install

Run this once in your terminal. Works globally across all Claude Code sessions.

claude mcp add llm-dojo --transport http https://llmdojo.dev/api/mcp -s user

Requires Claude Code CLI. Install at claude.ai/code

02 — What it does

Three tools are exposed. Claude uses them automatically when you ask LLM-related questions.

search_curriculum
search_curriculum("how to implement GRPO")

Semantic search over all 101 notebooks. Returns ranked results with descriptions, concept tags, and direct Colab links.

list_stages
list_stages()

Returns all 8 curriculum stages with notebook counts and descriptions. Useful for understanding what's covered.

get_notebook
get_notebook("62-grpo-reasoning")

Full details for a specific notebook — description, concepts, duration, Colab link.

03 — Example session

After connecting, Claude automatically queries the knowledge base when relevant:

you
How do I train a reasoning model like DeepSeek-R1?
claude code
→ calls search_curriculum("reasoning model DeepSeek GRPO")
→ returns: Notebook 62 — GRPO: Reasoning Model Training
   Group Relative Policy Optimization with verifiable rewards
   Colab: github.com/nishchaysinha/llm-dojo/blob/main/notebooks/62_grpo_reasoning.ipynb
you
What's the difference between DPO, ORPO, and KTO?
claude code
→ calls search_curriculum("DPO ORPO KTO comparison")
→ returns: Notebook 63 — Preference Algorithm Comparison
   Side-by-side DPO vs ORPO vs KTO vs IPO on same dataset

04 — Manual config (optional)

Or add directly to ~/.claude/settings.json for project-scoped use:

{
  "mcpServers": {
    "llm-dojo": {
      "type": "http",
      "url": "https://llmdojo.dev/api/mcp"
    }
  }
}

05 — Direct API

The search endpoint is also available as a plain HTTP API:

GET https://llmdojo.dev/api/search?q=LoRA+fine-tuning&k=5