Install | CriogenAI

Supported platforms#

	Linux x64	Linux ARM64	macOS Apple Silicon	macOS Intel	Windows x64
Python 3.9+ (one wheel per platform)	yes	yes	yes	yes	yes
Node.js 18+ (a prebuilt addon per platform)	yes	yes	yes	yes	yes
Rust 1.88+ (crates published when the code opens)	later	later	later	later	later
`criogenai` engine on the registry (Python / Node.js)	from 16 Nov	from 16 Nov	from 16 Nov	from 16 Nov	0.1.7 / 0.1.7
Until then: the old `criogen` package	0.1.3 / 0.1.3	0.1.3 / 0.1.3	0.1.3 / 0.1.3	0.1.2 / 0.1.3	–

Other platforms (Windows on ARM, Alpine/musl, 32-bit) are not supported yet.

Install#

pip install criogenai
# with Pydantic models as tool arguments and typed output:
pip install "criogenai[pydantic]"

The package is called criogenai; in code you import criogenai. The wheel contains the Rust engine, so nothing else is needed. It is typed (py.typed): mypy and your editor know every type.

npm install criogenai
# or, with another package manager (they all use the npm registry):
pnpm add criogenai
yarn add criogenai
bun add criogenai

npm installs the engine for your platform automatically (criogenai-linux-x64, -darwin-arm64, -windows-x64, …). Works from TypeScript and JavaScript, as CommonJS or ES modules.

The Rust crates are published when CriogenAI's code opens, at 10,000 installs (a Rust crate ships its source). Until then, use the Python or TypeScript package; they run the same Rust engine. This is what you will add then:

[dependencies]
criogenai-agent = { version = "0.1", features = ["sqlite"] }
criogenai-providers = "0.1"
criogenai-storage = { version = "0.1", features = ["sqlite"] }
serde_json = "1"
tokio = { version = "1", features = ["macros", "rt-multi-thread"] }

Crate	What it gives you
`criogenai-core`	The workflow engine: steps, edges, routers, state, pauses, retries
`criogenai-providers`	Messages, tools and models (OpenAI, Anthropic, Gemini, Ollama, vLLM, …)
`criogenai-agent`	Agents: tools, approvals, structured output, hooks, threads, memory
`criogenai-storage`	In memory, SQLite (`sqlite`), Postgres (`postgres`), encryption (`encryption`)
`criogenai-telemetry`	OpenTelemetry and Langfuse export
`criogenai-gateway`	The OpenAI-compatible LLM gateway (library and binary)

API keys#

Models read their keys from environment variables: OPENAI_API_KEY, ANTHROPIC_API_KEY, GEMINI_API_KEY, OPENROUTER_API_KEY. Self-hosted models (Ollama, vLLM, LM Studio) need none. Keys never appear in errors or logs: CriogenAI removes them from provider error messages.

Check the install#

python -c "import criogenai; print(criogenai.__version__, criogenai.ready_sync())"

node -e "import('criogenai').then(async (c) => console.log(await c.ready()))"

ready() loads the engine and reports the platform; it never throws. If the engine cannot load, see Troubleshooting.

Next#

Quickstart: an agent, a conversation and an approval in a few minutes.
Tutorial: build a support agent: a complete small app, step by step.