Installation

Largestack AI is a Beta Python framework. Requires Python >= 3.11 (3.11 / 3.12 / 3.13).

python -m pip install -U pip
pip install largestack

PyPI: https://pypi.org/project/largestack/
GitHub: https://github.com/Rivailabs/largestack

Verify the install

largestack --help
largestack version          # -> Largestack AI v1.1.1

import largestack
print(largestack.__version__)   # 1.1.1

Optional extras

Core install is provider-agnostic (it talks to OpenAI-compatible HTTP endpoints out of the box). Install extras for native SDKs and heavier features:

pip install "largestack[openai]"           # one extra
pip install "largestack[openai,rag,guard]" # several

Extra	Pulls in
`openai`	`openai`, `tiktoken`
`anthropic`	`anthropic`
`litellm`	`litellm`, `aiohttp` (multi-provider gateway)
`mcp`	`fastapi`, `uvicorn` (serve MCP)
`rag`	`sentence-transformers`, `faiss-cpu`, `qdrant-client`, `duckdb`, `beautifulsoup4`
`guard`	`presidio-analyzer`, `presidio-anonymizer` (PII guardrails)
`otel`	OpenTelemetry API + SDK
`office`	`openpyxl`, `python-pptx`, `pandas`
`tika`	`tika` (Apache Tika document parsing)
`postgres`	`psycopg`, `sqlalchemy`
`migrations`	`alembic`, `sqlalchemy`
`dev-server`	`fastapi`, `uvicorn`, `watchfiles` (hot-reload dev server)
`test`	`pytest`, `pytest-asyncio`, `pytest-timeout`, `hypothesis`, `respx`, `pytest-cov`
`docs`	`mkdocs-material`, `pymdown-extensions`, `mkdocstrings`
`all`	Everything above (except `docs`)
`dev`	`all` + `ruff`, `mypy`

pip install "largestack[all]"   # all optional features
pip install "largestack[dev]"   # contributor toolchain

If PyTorch tries to install large CUDA packages on Linux, prefer CPU wheels: PIP_EXTRA_INDEX_URL=https://download.pytorch.org/whl/cpu pip install "largestack[rag]"

Provider API keys

Set keys as environment variables using the pattern LARGESTACK_<PROVIDER>_API_KEY (provider name upper-cased). Never commit keys.

export LARGESTACK_OPENAI_API_KEY="sk-..."
export LARGESTACK_DEEPSEEK_API_KEY="sk-..."
export LARGESTACK_ANTHROPIC_API_KEY="sk-ant-..."

The model string selects the provider, e.g. Agent(name="r", llm="deepseek/deepseek-chat"). For local / OpenAI-compatible endpoints, see local LLMs and the provider support matrix.

You can develop and test without any key using TestModel / FunctionModel:

from largestack import Agent
from largestack.testing import TestModel

agent = Agent(name="r", llm="deepseek/deepseek-chat")
# with agent.override(model=TestModel(custom_output_text="canned")):
#     result = await agent.run("anything")   # no network, no key