Hands-On with Graflow: Run AI Workflows with Local LLMs in Google Colab
Want to try Graflow without installing anything? We've published a Google Colab notebook that walks you through building and running agentic workflows — right in your browser, completely free.
Try It Now
The notebook covers 9 topics with runnable examples — from your first graph to LLM-powered agents:
| # | Topic | What You'll Learn |
|---|---|---|
| 1 | Your First Graph | Define tasks with @task and wire them with >> — no upfront state schema needed |
| 2 | Parallel Execution | Fan-out / fan-in patterns using the | operator, plus dynamic task generation with chain() and parallel() |
| 3 | Data Sharing | Channel-based inter-task communication with concurrency-safe primitives (atomic_add, append, lock) and TypedChannel |
| 4 | Branching & Loops | Runtime control flow with next_iteration() for retries and next_task() for dynamic routing |
| 5 | Error Policies | Parallel group policies — Strict, Best-effort, At-least-N, and Critical |
| 6 | Exercise | Build a data analysis pipeline combining everything above |
| 7 | Checkpoint / Resume | Suspend and resume workflows with checkpoint() and CheckpointManager |
| 8 | Custom Task Handlers | Swappable execution strategies via TaskHandler subclassing |
| 9 | LLM Integration | Local LLMs with LiteLLM + Ollama, and Google ADK agents with tool calling |
Local LLMs: Privacy Without Compromise
The notebook includes examples using Gemma 4 (Google's recently released 4.5B-parameter multimodal model with 128K context) and other local models through Ollama. Two integration approaches are demonstrated:
-
inject_llm_client— Call any LLM directly from a task handler via LiteLLM, which provides a unified API across 100+ providers. Point it at a local Ollama endpoint and no API key is required. -
inject_llm_agent— Wrap a Google ADK agent as a Graflow fat node, complete with tool calling and the ReAct pattern.
Running models locally means sensitive data never leaves your environment — no API keys, no external calls, full control over your inference stack. This matters for enterprise workflows where compliance and data sovereignty are non-negotiable. And because Graflow is framework-agnostic, swapping between local and cloud models requires zero changes to your workflow code.
Get Involved
Graflow is Apache 2.0 licensed and open source. We welcome issues and pull requests.