Install#

TensorCode has two implementations with the same contracts: the Python package tensorcode (the reference) and its TypeScript port. Artifacts, experience files and session files move between them, so you can train in one language and serve in the other.

	Python	TypeScript
Package	`tensorcode`	`tensorcode`
Runtime	Python 3.11 or newer	Node.js 20.16 or newer, ESM only
Numerics	PyTorch, CPU or GPU	Built-in CPU autograd; separate restricted program WebGPU backend
Source	TensaCo/tensacode-py	TensaCo/tensacode-ts

Python (pip)#

Python 3.11 or newer:

python -m pip install 'tensorcode[tools]'

The core package has no dependencies, and importing it loads neither PyTorch nor the network. Choose the extras for the parts you use:

Extra	Adds
`tools`	Owned models, training and Hugging Face loading (PyTorch, Transformers)
`vec`	Vector operations and compiled-program training (PyTorch, NumPy, safetensors)
`local`	Adapter for a local multimodal Transformers model
`diffusion`	`tools` plus diffusers, for image decoders
`pretrained`	Alias of `tools`
`dev`	pytest, build, Pillow and PyArrow. The full test suite also needs `diffusion`

Check the install:

python -c "import tensorcode; print(tensorcode.__version__)"

To work on the library itself, clone it and install it in editable mode:

git clone https://github.com/TensaCo/tensacode-py
cd tensacode-py
python -m pip install -e '.[tools,diffusion,dev]'
python -m pytest -q

TypeScript (npm)#

Node.js 20.16 or newer:

npm install tensorcode

The package is ESM only (import, not require) and ships its own type declarations. Source compilation depends on typescript, installed automatically by npm. The optional peer @huggingface/transformers is needed only for integrations.LocalModel.

Check the install:

node --input-type=module -e "import { version } from 'tensorcode'; console.log(version)"

To work on the library itself:

git clone https://github.com/TensaCo/tensacode-ts
cd tensacode-ts
npm install          # also builds dist/
npm test             # program interop tests need python3 and ../python/src

Imports side by side#

The entry points match module for module. TypeScript uses camelCase for the API and keeps Python's snake_case for saved and reported data.

from tensorcode import trace, training
from tensorcode.ops.vec import Classify
from tensorcode.ops.vec.encode import VocabularyEncoder
from tensorcode.tools.investigator import Investigator

import { trace } from 'tensorcode';
import { Trainer, loadExperience } from 'tensorcode/training';
import { Classify, VocabularyEncoder } from 'tensorcode/ops/vec';
import { Investigator } from 'tensorcode/tools';

Compiler development setup#

The compiler guides describe the current SDK source, not a guarantee about the latest registry release. Use the editable Python checkout or built TypeScript checkout above. Python's compiler and reference runtime need no extras; training compiled parameters needs vec. For the TypeScript interoperability tests, place the Python checkout at ../python, as in the umbrella repository layout.

Start with compiled programs in Python or TypeScript. The parser-free browser core is available from tensorcode/program/runtime; the event Runtime class uses the full tensorcode/program entry.

Next steps#

Architecture overview: how operations, tools, tracing, training and artifacts fit together.
Quickstart: Python or TypeScript.
TypeScript parity with Python.