Installation

ThinkLess needs Python 3.10 or newer and runs on Linux, macOS and Windows.

Pick what you need

The core install is small: pydantic, httpx, rich and typer. Everything that brings in heavy dependencies is an extra.

You want	Install	Pulls in
Rules, tracing, the CLI, Jev or any System One server	pip install thinkless	nothing heavy
Hosted LLMs through OpenRouter, OpenAI, Ollama or vLLM	pip install "thinkless[openai]"	openai
Claude through the Anthropic SDK	pip install "thinkless[anthropic]"	anthropic
GLiNER 2.5	pip install "thinkless[gliner]"	gliner2, transformers<5, torch
Laya	pip install "thinkless[laya]"	laya, transformers<5, torch
A local LLM or any Hugging Face classifier	pip install "thinkless[local-llm]"	transformers<5, accelerate, torch
All local models	pip install "thinkless[local]"	the three above
OpenTelemetry export	pip install "thinkless[otel]"	opentelemetry-sdk and the OTLP exporter
The public dataset benchmarks	pip install "thinkless[bench]"	datasets
The LangGraph adapter	pip install "thinkless[langgraph]"	langgraph
The OpenAI Agents SDK adapter	pip install "thinkless[openai-agents]"	openai-agents
The decision server (thinkless serve)	pip install "thinkless[server]"	fastapi, uvicorn
MCP tools (thinkless mcp)	pip install "thinkless[mcp]"	mcp
Everything	pip install "thinkless[all]"	all of the above

Extras combine: pip install "thinkless[gliner,laya,openai]".

Local models: install PyTorch first

pip cannot choose the right PyTorch build for your GPU, so the local extras take whatever torch it finds, and on Windows and macOS that is the CPU build. Everything still works, only slower. For a GPU, install PyTorch first from its own index, then ThinkLess:

NVIDIA GPU (Linux or Windows). CUDA 13.0 wheels cover current cards, including the RTX 50 series:

pip install torch --index-url https://download.pytorch.org/whl/cu130
pip install "thinkless[local]"

For older drivers, pick the matching CUDA version on pytorch.org.

Apple Silicon. The default PyPI build includes Metal (MPS) support:

pip install "thinkless[local]"

CPU only.

pip install torch --index-url https://download.pytorch.org/whl/cpu
pip install "thinkless[local]"

Then check what you got:

thinkless doctor

doctor prints the accelerator torch can see. If an NVIDIA GPU is present but torch is a CPU build, it says so and prints the command to fix it.

Conda

environment.yml creates an environment with everything, including the CUDA 13.0 build of PyTorch:

conda env create -f environment.yml
conda activate thinkless

It installs ThinkLess from the checkout in editable mode, for development. To use the released package in a conda environment, create the environment with Python 3.12, install PyTorch as above, then pip install "thinkless[local]".

Where files go

What	Where	Change it with
Model weights	the Hugging Face cache, ~/.cache/huggingface	HF_HOME
Banking77 test file for the benchmarks	~/.cache/thinkless/datasets
Traces written by the CLI	./.thinkless/traces	THINKLESS_TRACE_DIR
Shadow logs	wherever Shadow(log=...) points
Benchmark output	./.thinkless/bench/<run>	--out
API keys	environment variables, or ./.env for the CLI

Models download on first use: about 0.8 GB for GLiNER base, 0.9 GB for Laya and 3.4 GB for Qwen3-1.7B. thinkless doctor shows which ones are already cached.

Offline and air-gapped machines

Download the weights once on a connected machine (running thinkless demo does it), copy the Hugging Face cache directory, and set HF_HUB_OFFLINE=1 on the offline machine. Providers also accept a local directory in place of a model id, for example GLiNER("/models/gliner2.5-base-v1").

Windows notes

• On Windows, ThinkLess downloads models one file at a time. That avoids a race in the Hugging Face cache's symlink handling (WinError 1314). Enabling Developer Mode lets the cache use symlinks and saves disk space.
• If a console shows garbled characters in tables, use Windows Terminal or set PYTHONIOENCODING=utf-8.

Checking an install

thinkless --version
thinkless doctor
python -c "import thinkless; print(thinkless.__version__)"

A run that needs no downloads, for a quick end-to-end check:

python -m thinkless demo --help