- Getting Started
- Option B — Existing project
Getting Started
Option B — Existing project
Add Avalanche to an existing Python project and run a narrow local workflow target.
Use this path when Avalanche should live alongside existing application code.
Requirements
- Python 3.11, 3.12, or 3.13
- uv
- An LLM provider API key or Codex subscription only when the workflow uses agent steps
Add Avalanche
uv add avalanche-ai
Add the skill (optional)
Install the authoring skill in the same project when you want a coding agent to help build workflows:
npx skills add Trampoline-AI/avalanche
Configure a provider
Configure credentials in the environment that runs ava dev or the local
operator, using the environment variables recognized by your LiteLLM provider.
Keep provider secrets out of workflow modules and source control.
The model identifier belongs in the workflow; provider secrets belong only in the runtime environment.
Set a provider-qualified model identifier on an agent step or in workflow
agent_defaults; a value on the individual step takes precedence. See
Configure models and providers for credential
and model-selection examples, and declare agent steps
for the step-level model contract.
Start with a narrow flow target
Put a workflow in a known module, then start the local operator against that module or a flow-only directory:
uv run ava dev --flows ./flow.py
Avoid starting ava dev or ava operator from a repository root containing
unrelated Python. Discovery imports candidate Python modules; narrow targets
make that boundary explicit.
Continue with workflow authoring or operator-managed runs.