Skip to content

Falcon

Falcon runs your TPU/GPU experiments for you: you describe a workload in a small YAML file, submit it with one CLI, and Falcon provisions the hardware, runs the job, and collects the outputs. The same falcon CLI captures profiles and runs analyzers over the results — you never touch Kubernetes, node pools, or object storage directly.

What you can do

  • Run experiments — submit a workload and let Falcon schedule it onto a TPU/GPU cluster.
  • Profile workloads — capture profile artifacts from a run and collect them when the job finishes.
  • Analyze results — run built-in or custom analyzer plugins over an artifact and read their reports.
  • Re-run past experiments — reproduce, clone, or tweak any prior run without hunting for the original YAML.

Two ways to use Falcon

With an agent (recommended for most users) — install the falcon-workflow skill and let your agent drive Falcon in natural language. → Using Falcon with an Agent

Directly via the CLI — run falcon commands yourself; best for scripting, CI, or fine-grained control. → Getting Started

Start here

I want to…Go to
Run my first experimentGetting Started
Understand the experiment lifecycleExperiment Lifecycle
Profile a workloadAnalyzer Workflow
Run / read analyzersAnalyzer Workflow

Deploying or operating Falcon? Those runbooks live in the Falcon repository operators docs.