dataflowr — CLI, API & MCP server for the Deep Learning DIY course

The Deep Learning DIY course teaches PyTorch from scratch — tensors, autodiff, CNNs, RNNs, Transformers, VAEs, and diffusion models — through hands-on notebooks. Course resources:

• dataflowr/notebooks — all practical notebooks (PyTorch fundamentals → diffusion models)
• dataflowr/gpu_llm_flash-attention — implement FlashAttention-2 from scratch using Triton
• dataflowr/llm_controlled-generation — structured generation, meta-generation, and self-correction for LLMs
• dataflowr/llm_efficiency - KV Cache and LoRA for minGPT
• dataflowr/transcripts — 318 concept notes extracted from lecture transcripts, with timestamped quotes and cross-references

This package exposes the course as a CLI, REST API, and MCP server so AI agents can navigate and teach it.

Quick start with Claude Code

Option 1 — Hosted server (no install needed): add a .mcp.json file at the root of your project pointing to the shared instance:

{
"mcpServers": {
"dataflowr": {
"type": "http",
"url": "https://dataflowr.paris.inria.fr/mcp"
}
}
}

Option 2 — Local server: run it yourself with uv (downloads on first use):

{
"mcpServers": {
"dataflowr": {
"type": "stdio",
"command": "uv",
"args": ["run", "--with", "dataflowr[mcp]", "python", "-m", "dataflowr.mcp_server"]
}
}
}

Claude Code picks this up automatically when you open the folder.

To pre-approve all dataflowr tools (no per-call prompts), also add .claude/settings.json:

{
"permissions": {
"allow": ["mcp__dataflowr"]
}
}

For a global allow (all projects), add the same to ~/.claude/settings.json.

---

Install

# with uv (recommended)
uv pip install dataflowr # CLI only
uv pip install dataflowr[api] # CLI + REST API
uv pip install dataflowr[mcp] # CLI + MCP server
uv pip install dataflowr[all] # everything# with pip
pip install dataflowr
pip install dataflowr[mcp]

Or from source:

git clone https://github.com/dataflowr/dataflowr-tools
cd dataflowr-tools
uv pip install -e ".[all]"

---

CLI

# Course overview
dataflowr info
# List all modules
dataflowr modules list
# Filter by session, tag, or GPU requirement
dataflowr modules list --session 7
dataflowr modules list --tag attention
dataflowr modules list --gpu
# Full module details (with notebook links)
dataflowr module 12
# Fetch notebook content from GitHub
dataflowr notebook 12 # practical (default)
dataflowr notebook 12 --kind intro
dataflowr notebook 12 --no-code # markdown only# Fetch course website page text (raw markdown from dataflowr/website)
dataflowr page 12
# Fetch lecture slides (from dataflowr/slides)
dataflowr slides 12
# Fetch quiz questions (from dataflowr/quiz)
dataflowr quiz 2a
dataflowr quiz 3
# Browse transcript knowledge base (318 concept notes from lectures)
dataflowr transcripts search "backprop"
dataflowr transcripts get "training loop"# Compare catalog against website + slides repos
dataflowr sync
# Search by keyword
dataflowr search "attention transformer"
dataflowr search "generative"# Sessions
dataflowr sessions list
dataflowr sessions get 7
# Homeworks
dataflowr homeworks list
dataflowr homeworks get 1
# JSON output (pipe-friendly)
dataflowr modules list --json | jq '.[] | select(.session == 9)'
dataflowr module 18b --json

---

REST API

uvicorn dataflowr.api:app --reload
# → http://localhost:8000# → http://localhost:8000/docs (Swagger UI)

Endpoints:

Method	Path	Description
GET	/	Course overview
GET	/modules	List all modules (?session=, ?tag=, ?gpu=)
GET	/modules/{id}	Get module by ID
GET	/modules/{id}/notebooks	Get notebooks for a module (?kind=)
GET	/modules/{id}/notebooks/{kind}/content	Fetch notebook cells from GitHub (?include_code=)
GET	/modules/{id}/slides	Fetch lecture slide content from dataflowr/slides
GET	/modules/{id}/quiz	Fetch quiz questions from dataflowr/quiz
GET	/modules/{id}/page	Fetch module source markdown from dataflowr/website
GET	/catalog/sync	Compare catalog against website + slides repos
GET	/sessions	List all sessions
GET	/sessions/{n}	Get session with modules
GET	/homeworks	List all homeworks
GET	/homeworks/{id}	Get homework by ID
GET	/search?q=...	Search modules
GET	/transcripts/search?q=...	Search transcript concept notes
GET	/transcripts/{concept}	Get a transcript concept note

Examples:

curl http://localhost:8000/modules/12
curl http://localhost:8000/sessions/7
curl http://localhost:8000/search?q=diffusion
curl "http://localhost:8000/modules?session=9&gpu=true"
curl "http://localhost:8000/modules?tag=attention"
curl "http://localhost:8000/modules/12/notebooks/practical/content?include_code=false"
curl http://localhost:8000/modules/12/page
curl http://localhost:8000/modules/2a/quiz
curl http://localhost:8000/modules/3/quiz
curl "http://localhost:8000/transcripts/search?q=backprop"
curl http://localhost:8000/transcripts/training%20loop

---

MCP Server (for AI agents)

Makes the course natively available to Claude, Cursor, VS Code, and other MCP-compatible agents. Built on the official MCP Python SDK (FastMCP).

Stdio transport (local — Claude Desktop, Cursor, VS Code, Claude Code)

python -m dataflowr.mcp_server

HTTP transport (remote / shared deployments)

python -m dataflowr.mcp_server --http
# → POST http://localhost:8001/mcp (or $PORT if set)

---

Client configuration

Claude Code (VSCode extension or CLI)

Add a .mcp.json file at the root of your project (homework repo, notebook folder, etc.):

{
"mcpServers": {
"dataflowr": {
"type": "stdio",
"command": "uv",
"args": ["run", "--with", "dataflowr[mcp]", "python", "-m", "dataflowr.mcp_server"]
}
}
}

Claude Code picks this up automatically when you open the folder. No global install needed — uv downloads dataflowr[mcp] on first use.

To pre-approve all dataflowr tools (no per-call prompts), also add .claude/settings.json:

{
"permissions": {
"allow": ["mcp__dataflowr"]
}
}

Or register globally (available in every project):

claude mcp add --scope user dataflowr -- uv run --with dataflowr[mcp] python -m dataflowr.mcp_server

Claude Desktop

Edit ~/.claude/claude_desktop_config.json (macOS/Linux) or %APPDATA%\Claude\claude_desktop_config.json (Windows):

{
"mcpServers": {
"dataflowr": {
"command": "python",
"args": ["-m", "dataflowr.mcp_server"]
}
}
}

With uv (no global install needed):

{
"mcpServers": {
"dataflowr": {
"command": "uv",
"args": ["run", "--with", "dataflowr[mcp]", "python", "-m", "dataflowr.mcp_server"]
}
}
}

Cursor

Edit .cursor/mcp.json at the root of your project (or ~/.cursor/mcp.json globally):

{
"mcpServers": {
"dataflowr": {
"command": "python",
"args": ["-m", "dataflowr.mcp_server"]
}
}
}

VS Code

Edit .vscode/mcp.json at the root of your project:

{
"servers": {
"dataflowr": {
"type": "stdio",
"command": "python",
"args": ["-m", "dataflowr.mcp_server"]
}
}
}

Remote / HTTP (self-hosted)

If running your own instance with --http, point clients at the URL:

{
"mcpServers": {
"dataflowr": {
"type": "http",
"url": "http://localhost:8001/mcp"
}
}
}

---

Recommended workflow

1. search_modules "attention" → find relevant modules by keyword
2. get_module "12" → full details, notebook links, prerequisites
3. get_page_content "12" → read the lecture notes
4. get_notebook_content "12" → work through the exercises
5. get_quiz_content "12" → self-test your understanding
6. search_transcripts "attention" → find concept notes from lecture transcripts
7. get_transcript_note "training loop" → read timestamped quotes and cross-references

For a personalised study plan, use the learning_path prompt with a target module.

---

Tools exposed to the agent

Tool	Description
list_modules	List modules, filterable by session / tag / GPU
get_module	Full details for a module
search_modules	Keyword search across titles, descriptions, and tags
list_sessions	List all sessions
get_session	Session + all module content
get_notebook_url	GitHub/Colab links for a notebook
list_homeworks	All homeworks
get_homework	Full details for a homework
get_slide_content	Fetch lecture slides from dataflowr/slides
get_quiz_content	Fetch quiz questions from dataflowr/quiz
check_quiz_answer	Validate a student's quiz answer
get_notebook_content	Fetch actual notebook cells from GitHub
get_notebook_exercises	Fetch only exercise prompts + skeleton code
get_page_content	Fetch module source markdown from dataflowr/website
get_course_overview	Full course structure as context
get_prerequisites	Prerequisite modules for a given module
suggest_next	What to study after completing a module
sync_catalog	Compare catalog against website + slides repos
search_transcripts	Fuzzy search 318 concept notes from lecture transcripts
get_transcript_note	Fetch full content of a transcript concept note

Prompts exposed to the agent

Prompt	Arguments	Description
explain_module	module_id	Tutoring session — Socratic explanation of a module
quiz_student	module_id	Interactive quiz, one question at a time
debug_help	module_id	Guided debugging help for a practical notebook
learning_path	target_module_id, known_modules	Personalised prerequisite chain to a target module

---

Example questions

Once connected, an agent can answer questions like:

• "What should I study before tackling diffusion models?"
• "Give me the Colab link for the microGPT notebook."
• "Which session covers attention mechanisms?"
• "What are all the generative modeling modules?"
• "Show me the Flash Attention homework tasks."
• "Quiz me on Module 3 — loss functions."
• "I'm stuck on the backprop exercise in Module 2b. Help me debug it."
• "Build me a learning path to Module 18b (diffusion models) starting from scratch."
• "What does the professor say about the training loop? Show me the transcript quotes."

---

Python API

fromdataflowrimportCOURSE# Get a modulemodule=COURSE.get_module("12")
print(module.title) # "Attention and Transformers"print(module.description)
print(module.notebooks)
# Searchresults=COURSE.search("attention")
# Get a sessionmodules=COURSE.get_session_modules(7)
# Navigate the full catalogformoduleinCOURSE.modules.values():
ifmodule.requires_gpu:
print(f"Module {module.id}: {module.title}")

---

Design principles

• Content only, no execution. The package exposes the course structure and links. Running notebooks stays in the student's hands.
• Agent-friendly. All outputs are text-first. The MCP server renders markdown so agents can use it directly in responses.
• No external dependencies for core. The catalog, models, and CLI work with only pydantic, typer, and rich. The API needs fastapi; the MCP server needs mcp.
• Single source of truth.catalog.py is the only place that needs updating when the course evolves.