Scratch-like drag-and-drop visual programming for AI/ML engineering. Have an idea, refine with drag & drop, and receive functional code immediately.
Built for anyone to understand, inspect, trust, and build with AI.
AI Blocks is an open source visual programming environment for AI/ML pipelines. Instead of starting with a blank notebook, you compose models, data, agents, and evaluation steps as blocks — each one a small, typed, inspectable unit — and the assembler turns the graph into runnable Python.
The goal is simple: lower the floor for building real ML systems without hiding what's underneath. Every block maps to code you can read, copy, and ship.
- Learn while building. Each block shows its inputs, outputs, and the exact code it produces.
- Compose, don't copy-paste. Pipelines are graphs, not fragile notebooks.
- Typed ports. Connections are validated so you don't wire a tensor into a string.
- Real code, not a toy. The assembler emits plain Python you can run, commit, and extend.
Blocks live in packages/block-schemas and the Python runtime in aiblocks/. Current categories:
| Category | What's inside |
|---|---|
transformers-llms |
LLM calls, tokenizers, chat formatting |
agents |
Tool-using agents, planning loops |
neural-networks |
Layers, optimizers, training loops |
classical-ml |
Tabular models, train/test splits, metrics |
embeddings-retrieval |
Embeddings, vector search, RAG |
prompt-engineering |
Prompt templates, few-shot, output parsing |
vision |
Image IO, preprocessing, detection, pose |
audio-speech |
ASR, TTS, voice cloning |
text-nlp |
Classification, sentiment, NER |
data-io / data-processing |
CSV, JSON, batching, cleaning |
training / fine-tuning / distillation |
Training, LoRA, KD |
evaluation / monitoring / experiment-tracking |
Metrics, logging, tracing |
control-flow / utilities |
Branches, loops, helpers |
pnpm install
pnpm devThen open the web app and pick a template, or start from an empty canvas.
pnpm build # build all packages
pnpm validate-blocks # validate block schemas
pnpm gen-block # scaffold a new blockRequires Node >= 20 and pnpm >= 9.
apps/web Vite + React editor (landing, canvas, docs)
packages/block-schemas Typed block definitions (the block library)
packages/assembler Graph → Python code assembler
packages/core Shared core types and runtime glue
packages/ui-components Editor UI primitives (toolbar, canvas, panels)
aiblocks/ Python runtime implementations for each block
extensions/ VS Code extension
benchmark/ Pipeline benchmarks
docs/plans Architecture, assembler, and embedding plans
docs/screenshots README and marketing screenshots
This is the main way to help. Every new block unlocks new pipelines for everyone.
-
Scaffold it
pnpm gen-block
-
Define the schema in
packages/block-schemas/src/blocks/<category>.ts— name, description, typed inputs/outputs, parameters. -
Implement the runtime in
aiblocks/<category>.py— a plain Python function the assembler calls. -
Validate
pnpm validate-blocks
-
Open a PR. Include a minimal example pipeline that uses the block, and a screenshot if it's visual.
Good first blocks: anything wrapping a single well-known model, dataset loader, or metric you already use. See docs/plans/blocks.txt for the running wishlist and docs/plans/architecture.md for the design.
docs/plans/architecture.md— overall architecturedocs/plans/assembler-plan.md— graph → code assemblerdocs/plans/embedding-plan.md— embedding/retrieval designdocs/plans/blocks.txt— block wishlist
- Issues: bug reports, block requests, and design discussion are all welcome.
- Pull requests: new blocks, new templates, docs fixes — all appreciated.
- Templates: the landing page ships with starter templates (Auto Research Agent, GPT-4o Chat, March Madness, OpenClaw, Polymarket, Pushup Pose, ElevenLabs Translator, AlphaGo). Adding a template is a great first contribution.
Open source. See LICENSE for details.