CanyonOS

CanyonOS

CanyonOS is an open source platform for multi-agent applications written in plain Python. Your agents do the work, with no SDK, no decorators and no tracing code, and CanyonOS runs everything around them, locally or on your own infrastructure. You describe how your workflow runs in a workflow config file and run canyonos deploy.

Core features

With CanyonOS Without CanyonOS
Orchestration One-Click deploy Set up Kubernetes or other infrastructure
Observability Built in, exports to OpenTelemetry Add and wire up an observability platform
Async execution Built in, each agent scales alone Add Ray or another scaling framework
Cost Per workflow and per agent, within the limits you set Calculations and spreadsheets

What CanyonOS offers has the full list, and what CanyonOS works with.

How we think about workflows

A workflow is a program that delivers an outcome, and each agent is a specialist it trusts with one step.

Your project folder is the workflow, and one lead agent receives each request and calls the others. Each agent is a class in its own file, and one agent calls another as if it were a local function; CanyonOS carries the call to the other agent's container. Your code says what the application does, and the config says how it runs: how many replicas, what resources, which models.

Give each job its own agent, so a busy search agent can grow to ten copies while the rest stay at one. A few well-chosen agents are easier to run than many small ones. Workflow design has the full approach.

Try it

With CanyonOS installed, run these in an empty folder:

canyonos new-app     # writes a starter workflow with two small agents
canyonos deploy      # runs it

Start from scratch takes it from there, step by step.

Contribute

CanyonOS is open source under the AGPL-3.0, and the code is on GitHub. Issues and pull requests are welcome. Contributing covers setting up a checkout, running the tests and how pull requests are reviewed. CanyonOS is maintained by Canyon Code.

Where to go next