multicloud-cli

mctl connects your AWS account to Multicloud. Multicloud then starts machines in your account, and AWS bills you directly. You give Multicloud no password and no access key: you create one IAM role with the public Terraform module multicloud/multicloud/aws.

The package is not on PyPI yet, so these commands do not work today. When it is published, install it with:

pipx install multicloud-cli        # or: uv tool install multicloud-cli

If mctl is not found after the install, run uv tool update-shell (or pipx ensurepath) and open a new shell.

Connect an AWS account

You need to be an owner of your team on multicloud.io, an AWS login that can create IAM roles, and OpenTofu or Terraform 1.5 or later.

mctl auth login                       # opens your browser; sign in as you do on multicloud.io
mkdir multicloud-aws && cd multicloud-aws
mctl cloud-account create --cloud AWS --account 123456789012 --region us-east-1
mctl cloud-account apply 7            # init, plan, asks you, apply, then waits for "Ready"

Keep the folder: it holds the Terraform state of the role.

To run OpenTofu or Terraform yourself, get a one-time connect code instead of apply:

mctl cloud-account token 7            # prints one "export TF_VAR_multicloud_connect_token=..." line
tofu init && tofu apply
mctl cloud-account verify 7 --wait

Commands

The connect code reaches Terraform only through the environment of the apply process. mctl never prints a key or a token, and only mctl cloud-account token prints a connect code.

Sign in

mctl auth login opens your browser on auth.multicloud.io (authorization code with PKCE, on a temporary port of 127.0.0.1). mctl auth login --no-browser uses the device flow instead: it prints a web address and a code, and you sign in on any device. Both give a short-lived access token (15 minutes) and a refresh token. mctl keeps them in the OS key store, else in ~/.config/multicloud/credentials (mode 0600), and refreshes the access token by itself. A session ends after at most 10 hours; then run mctl auth login again.

• Organization. The sign-in holds the organizations you belong to. With one, no --org is needed. With several, name one: --org <team-id> or MULTICLOUD_ORG.
• Order of credentials. MULTICLOUD_API_KEY (for CI), then the SSO sign-in, then a stored API key.
• Roles. Pool reads need membership. Everything that changes a pool, and every cloud-account command, needs the owner role of the team (an API key with the matching scope works too).
• Logout. mctl auth logout ends the session at Keycloak and deletes the local copy. Because mctl shares your Keycloak browser session, it signs you out of Keycloak in your browser too.
• Settings. MULTICLOUD_ISSUER_URL (or issuer in ~/.config/multicloud/cli.json) overrides the Keycloak realm URL https://auth.multicloud.io/realms/multicloud. Like the API URL it must be https://.
• Other environments. The defaults point to production (auth.multicloud.io, api.multicloud.io). For another environment, set MULTICLOUD_ISSUER_URL (for example https://auth.example.com/realms/multicloud) and MULTICLOUD_API_URL (for example https://api.example.com) to that environment's hosts before mctl auth login. Set both: the sign-in server (issuer) and the API must belong to the same environment, or the API rejects the token.
• API key (CI, scripts). mctl auth login --api-key --org <team-id>, or MULTICLOUD_API_KEY in the environment.

Manage CI pools

A pool is a group of build machines that Multicloud starts for you and stops when they are idle. With an API key, reading needs Read access, create and update need Write, and delete, hibernate and wake need Admin. With an SSO sign-in, reading needs team membership and the rest needs the owner role. Every command takes --org <team-id> (also before the command: mctl --org acme pool list) and, where it makes sense, --json for the raw answer. pool list and cloud-account list take -q (only the IDs) and --no-headers. Data goes to stdout; progress and questions go to stderr. <pool> is the number from mctl pool list or the exact pool name.

mctl pool classes                                  # the machine sizes you can ask for
mctl pool create ci --github-org my-gh-org --max-cost-per-hour 2 --classes cpu-4 \
  --clouds '[{"type":"AWS","regions":["us-east-1"]}]'
mctl pool list
mctl pool hibernate ci                             # pause: no machines, no cost, the name stays
mctl pool wake ci --wait --timeout 15m             # exit 0 ready, 5 timeout, 6 pool deleted, 1 other error

Find servers and prices

The catalog lists the servers of AWS, Azure and Google Cloud with prices. An API key needs the Catalog access (catalog scope); MULTICLOUD_API_KEY needs no --org. Filter names are the API names: --min-vcpus and --min_vcpus both work. Data goes to stdout; the page line (Page 1 of 3, 42 results.) goes to stderr.

mctl catalog search --provider aws --min-vcpus 8 --max-price-per-hour 0.5 --sort-by vcpus
mctl catalog search --has-accelerator --accelerator-name 'A100.*' --price-type SPOT -q   # only the IDs
mctl catalog show 11
mctl catalog compare 11 12 13
mctl catalog locations --provider aws
mctl catalog providers
mctl catalog field-values                          # the values the filters accept

Every command takes --json. See docs/reference.md for all options.

Use it from an AI agent

The skill multicloud-connect-aws in multicloud/skills teaches an agent such as Claude Code to run these commands. The agent asks you before each apply and never sees the key or the code.

More

Every command shows its options with --help or -h, for example mctl pool create -h. The exit codes, shell completion and all options are in the command reference: 0 ok, 1 error, 2 usage, 3 sign-in or access, 4 not found, 5 still pending or wait timeout, 6 pool gone.

License: Apache-2.0.