Install

Install quickstarted and check what your machine can enforce.

Agent mode is what you came for, so install a vendor SDK with it:

pip install "quickstarted[claude]"       # Anthropic
pip install "quickstarted[openai]"       # OpenAI
pip install "quickstarted[gemini]"       # Google
pip install "quickstarted[all-agents]"   # all three

Python 3.9 or newer. The only runtime dependency the harness itself adds is PyYAML; the rest is the SDK you chose.

pip install quickstarted

The bare install runs replay mode only, which needs no model and no key. That is the right install for a CI job that just checks the documented commands still work.

Check your machine

quickstarted doctor

quickstarted doctor
  backends available: docker, seatbelt, local
  auto would choose:  docker
  price book loaded:  no (token counts only)
  QUICKSTARTED_ANTHROPIC_API_KEY: set

Tasks execute commands that a model wrote after reading somebody else's documentation, which is untrusted code by any reasonable definition. The backend decides what those commands can touch:

Backend	Available when	Enforced
docker	a Docker daemon is running	yes
seatbelt	macOS	yes
local	always	no

If doctor reports local only, install Docker before you point this at a project you did not write. quickstarted run will refuse to use local until you pass --allow-unenforced. See Sandboxing.

Set an API key

Keys are read from QUICKSTARTED_* names first:

export QUICKSTARTED_ANTHROPIC_API_KEY=sk-ant-...

The vendor-standard names (ANTHROPIC_API_KEY, OPENAI_API_KEY, GOOGLE_API_KEY) work as a fallback, which is what CI usually sets. The prefixed names exist so a key can live in your shell without other tooling on the same machine finding it and billing against it. Nothing in the sandbox ever sees either name; the executor builds a scrubbed environment, and a test asserts that no key reaches a command.

Next: your first run.