CanyonOS
CanyonOS deploys and scales plain-Python agent workflows on Docker, AWS, or Kubernetes, with per-agent observability and governance. It is for developers and platform teams running agentic applications who want to manage deployment and scaling in their own environment.
Who it's for
- Python developers building agentic workflows who want to deploy agents without a CanyonOS SDK, decorators, or graph definition.
- Platform and infrastructure teams deploying agentic applications to Docker, AWS, or an existing Kubernetes cluster.
- Teams that need per-agent scaling, telemetry exports, or policy-based model and spend rules for agent workflows.
Not the right fit if…
- Teams expecting an established record at large production scale: CanyonOS is at version 0.2 and the page says it has not run at large production scale.
- Teams that need to keep agents in their current LangChain or LangGraph form without adapting them: the page says agents must be rewritten into plain Python, though an optional helper can convert a copy for review.
How it fits your workflow
Write agents in plain Python, with one file per agent. An optional helper can convert a copy of a LangChain or LangGraph project for review.
Run the quickstart locally with the CLI, or deploy workflows to AWS or Kubernetes from the same CLI; CanyonOS also runs on Docker. Bring a model provider and keys when deploying your own workflow.
Scale agents individually, set per-agent spend and model rules in a policy file, and export OpenTelemetry-compatible traces, metrics, and logs to supported observability tools.
Pricing
Free: CanyonOS is free and open source under AGPL-3.0. The page says paid services around it may be offered later.
Open source
Free
- · CanyonOS is open source under AGPL-3.0.
- · Users can download, run, and modify CanyonOS.
- · If modified and offered to others over a network, AGPL requires making the source of those changes available to them.
Prices checked on Nov 13, 2026 from the vendor's site. They can change; confirm before you buy.
Key features
- Plain Python workflows
- Write agents in plain Python without a CanyonOS SDK, decorators, or a graph definition. The page says agents are organized one file per agent.
- One-command deployment
- Deploy to an AWS account or Kubernetes cluster from the same CLI. CanyonOS also runs on Docker.
- Independent scaling
- Scale an individual agent rather than scaling the entire application.
- Built-in observability
- Get per-workflow, per-agent, and per-query insights without adding tracing code. CanyonOS exports OpenTelemetry-compatible traces, metrics, and logs.
- Policy-based governance
- Set per-agent spend and model rules in a policy file and change those rules without redeploying.
- Model and environment flexibility
- Run on a laptop, AWS, or Kubernetes and use providers with an OpenAI-compatible API, including common servers for self-hosted models.
Works with
- OpenAI-compatible model providers
- Datadog, Grafana, Prometheus, Langfuse, and Arize Phoenix
- Docker, AWS, and Kubernetes
- LangChain and LangGraph
Limitations to know
- CanyonOS is at version 0.2, its first public release, and the page says it has not run at large production scale.
- Agents need to be written in plain Python, one file per agent; adapting existing LangChain or LangGraph projects may require the optional conversion helper and user review.
- Deploying a user's own workflow requires bringing a model provider and keys; the no-key claim applies to the laptop Hello World quickstart.
- The page says CLI usage telemetry is sent by default and can be disabled with one setting.
- Under AGPL-3.0, users who modify CanyonOS and offer it to others over a network must make the source of those changes available.
Alternatives to consider
-
AI Elements by Vercel
Choose it if you want the agent infrastructure described as including sandboxed VMs, durable orchestration, and model gateway access rather than CanyonOS's plain-Python workflow deployment to AWS or Kubernetes.
Getting started
curl -fsSL https://raw.githubusercontent.com/CanyonCodeCoreAI/canyonos/main/packages/cli/install.sh | sh
Docker is supported for running CanyonOS; Kubernetes is optional. The laptop Hello World quickstart needs no account or API key. Deploying your own workflow requires a model provider with an OpenAI-compatible API and your own keys.
- Install CanyonOS using the supplied install script.
- Run `canyonos quickstart` to try the Hello World quickstart on a laptop.
- Write agents in plain Python, one file per agent; optionally use the helper to convert a copy of a LangChain or LangGraph project for review.
- Bring your model provider and keys, then deploy to AWS or Kubernetes with the CLI, or run on Docker.
Frequently asked questions
Do I need an account to run CanyonOS?
No. You can install it and run the Hello World quickstart on your laptop without an account or API key. Bring your own model provider when deploying your own workflow.
Do I need Kubernetes?
No. CanyonOS runs on Docker. If you already use Kubernetes, it can deploy there with an operator and a Helm chart.
Do I have to rewrite my agents?
Yes. Agents need to be written in plain Python, one file per agent. An optional helper converts a copy of a LangChain or LangGraph project for you to check.
Which model providers does CanyonOS support?
Any provider with an OpenAI-compatible API. The page also notes that common servers for models hosted by users offer such APIs.
Can I keep my existing observability tools?
Yes. CanyonOS exports traces and metrics to existing tools and adds cost and queue time per agent.
Is CanyonOS production-ready?
The supplied page describes version 0.2 as its first public release and an active project with weekly releases. It says CanyonOS has not yet run at large production scale and that current limits are listed in the README.
What does AGPL-3.0 mean for my company?
You can download, run, and change CanyonOS for free. If you modify it and offer it to others over a network, AGPL requires you to make the source of those changes available to them.
What information does the CLI send home?
The CLI sends an anonymous install ID, the command, whether it worked, and the CLI version. The page says it does not send code, prompts, or model data, and that telemetry can be turned off with one setting.