Skip to content
Back to Home

Introducing the SaladCloud CLI: From Your Terminal to CI/CD and AI Agents

Deploy, automate, and operate SaladCloud from the command line, with repeatable workflows for developers, CI/CD pipelines, and AI coding agents.

Maksim Gorkii
SaladCloudTutorials

We know many of our customers and developers have been waiting for this, and we are excited to introduce the SaladCloud CLI: command-line tool that brings SaladCloud into the workflows you already use to build and ship your applications.

You can deploy a container group, update an image, check GPU prices, inspect instances, and follow logs from your terminal. The same interface works in a deployment script, a CI/CD pipeline, or a coding agent.

That last part is where it gets especially useful. You can go from a configuration you tested locally to an automated deployment without rebuilding the integration around a different tool. And when an agent helps with the application, it can also work with the infrastructure through commands you can inspect and understand.

The CLI is available in beta today. Start with the CLI documentation, or find the source and releases on GitHub.

Less Integration Code Between Your Application and Its Infrastructure

The Portal is a good place to explore SaladCloud and launch a workload. Once that workload becomes part of a product, you usually want to repeat the same steps: deploy a new image, change a resource setting, check whether instances are ready, or collect logs when something goes wrong.

The CLI reduces the custom API or SDK code needed to automate those steps and gives you a common interface for that work. You can start with an individual command and keep using it when the process moves into a script.

For teams, that means fewer one-off deployment helpers to maintain. For someone testing a new model or worker image, it means less time moving between tools just to find out what happened.

Bring Deployments into CI/CD

A useful deployment pipeline should be able to build an image, publish it to a registry, update the workload, and verify the result. Your existing CI system still handles the build and registry push. The SaladCloud CLI handles the SaladCloud side.

For example, after your pipeline publishes an image and sets IMAGE to its tag, it can update an existing container group with:

saladcloud container-groups update my-app --image "$IMAGE" --timeout 5m -o json

For private images, supply registry credentials again under container.registry_authentication in a -f patch file generated from CI secrets. Keep that file out of version control.

Provide SALAD_API_KEY, SALAD_ORGANIZATION, and SALAD_PROJECT through the CI environment, with the key stored as a secret. There is no interactive login step for the runner. JSON output and documented exit codes let the pipeline react to results without trying to interpret a human-readable table.

This fits naturally into GitHub Actions, GitLab CI, or another system that can run a command-line tool. You choose when to deploy and what approval gates to use; the CLI gives the pipeline a consistent way to carry out the change.

A successful update is not the same as an application being ready. Check instance readiness and test the endpoint before treating a deployment as complete. Updating an image restarts the group's instances, so use a separate group and an explicit traffic switch when your application needs to avoid that interruption.

The usage and automation guide includes complete examples you can adapt to your pipeline.

Make Deployments Repeatable

A command is convenient for a quick experiment. A configuration file is easier to review and reuse.

The CLI accepts container-group definitions in YAML or JSON. You can keep a non-secret definition alongside your application, review resource changes in a pull request, and use it as the starting point for a new deployment. Instead of reconstructing settings from memory, you have a record of what you intended to run.

You can also create a group from an existing one. That makes it practical to test a different model image or resource configuration next to your current workload without first changing the working version. Copies do not include private registry credentials, so supply those again when creating the new group.

For benchmarking, this is particularly useful: repeat the deployment process across different GPU choices, collect results, and compare the cost of doing the actual work. The CLI does not run the benchmark for you, but it removes some of the repetitive setup around it. Keep private registry credentials and other secrets out of committed configuration files.

Give Your Coding Agent a Way to Operate SaladCloud

Coding agents are already useful for writing application code, building containers, and debugging failures. The next step is giving them a reliable way to work with the environment where that code runs.

From your project directory, run:

saladcloud agent init

This adds SaladCloud guidance to the project's agent instruction file, such as AGENTS.md or CLAUDE.md. It does not put your credentials into the repository.

There are three complementary pieces:

  • The CLI performs operations: inspecting resources, deploying an image, changing scale, or reading logs.
  • SaladCloud skills provide workflows: how to approach a task and check its result.
  • The documentation MCP server provides reference material: current SaladCloud documentation the agent can consult.

The documentation MCP server is not a deployment server. The CLI is what carries out the operational commands.

That combination lets you give an agent a concrete task, such as:

Inspect the failing container group in my staging project, review its logs, and explain what needs to change. Ask before modifying the deployment.

Keep API keys out of prompts, specify the organization and project, and set approval boundaries for billable or destructive actions. Those boundaries belong in your agent setup; they are not an automatic sandbox enforced by the CLI.

See using the CLI with AI agents and Build with AI Agents for the setup.

More Than Deployments

The same tool also helps with day-to-day operations. You can inspect GPU pricing and estimated capacity before a test, follow application logs and system events, and open a shell or copy files when investigating a running instance.

It also covers job queues, transcription, and AI Gateway model discovery and key testing. That makes it useful for smaller tasks too: submitting work from a script, transcribing a local recording, or checking a service without building a separate integration first. Capacity estimates are a planning aid, not a reservation.

The point is not to replace every tool you already use. It is to make SaladCloud easier to connect to those tools.

Get Started in a Few Commands

You will need a SaladCloud account, an organization with credits, and an API key. The Portal still handles account setup and billing.

On macOS or Linux:

curl -fsSL https://raw.githubusercontent.com/SaladTechnologies/salad-cloud-cli/main/install.sh | sh

On Windows, in PowerShell:

irm https://raw.githubusercontent.com/SaladTechnologies/salad-cloud-cli/main/install.ps1 | iex

Then check the installation and log in from your own terminal:

saladcloud version
saladcloud login
saladcloud doctor
saladcloud gpus list
saladcloud container-groups list

The installers download a release and verify its checksums. Follow their PATH instructions if your terminal cannot find saladcloud.

Because the CLI is in beta, pin the version in automation and check the release notes before upgrading. Use the installed command's --help for the flags available in your version.

Start with one task you currently do by hand. Put it in a script, add it to your pipeline, or give it to an agent with a clear scope. That is the main benefit of the CLI: the work you do once can become a workflow you keep using.

Read the SaladCloud CLI documentation.