Drive Excloud from an AI Coding Agent

The fastest way to operate Excloud is to let an AI coding agent drive the exc CLI for you. The Excloud team publishes a bundle of agent skills — short, declarative instruction packs — that teach agents how to use exc correctly: which flags exist, when to confirm before doing something destructive, how to chain CLI calls into a workflow.

Today the bundle is tested against:

• Claude Code
• pi coding agent

…and any other agent runtime that supports the Skill spec (npx skills add ...).

What’s in the bundle#

Skills are user-invocable building blocks that an agent will reach for when the task fits. The bundle currently ships:

More skills are added as the platform grows (object storage, secrets, Terraform). Re-run the install command to pick them up.

Install#

The Excloud installer adds the skills for you automatically. If you ran:

curl -fsSL https://excloud.dev/install.sh | bash

…the skills are already installed globally for any agent on your machine.

To install (or refresh) manually:

npx skills add https://git.excloud.dev/excloud-dev/excloud-skills.git --all --global

--global installs the skills under ~/.skills/ so every agent on the machine sees them. Drop the flag to install per-project under ./.skills/.

Using them with Claude Code#

Once installed, Claude Code surfaces the skills automatically. You can:

• Type a slash command — /excloud-cli list my VMs in zone 1.
• Ask in natural language — “Create a 4-core Ubuntu VM with a reserved public IP and put it behind a security group that allows HTTP.” Claude will pick up the excloud-cli skill and start chaining exc calls.

The skill is read-mostly until it needs to spend money or delete something. For anything destructive (terminate, delete, release, etc.) the skill is wired to confirm before acting.

Using them with pi#

pi is a CLI-first coding agent that uses the same skill format. Install pi:

npm install -g @mariozechner/pi-coding-agent

Then run:

pi "spin up a t1a.micro in zone 1 with my SSH key and an ephemeral public IP, show me the IP when it's running"

Pi reads ~/.skills/excloud-cli, plans the exc calls, and runs them, surfacing each command and its output as it goes.

A typical session#

This is a real Claude Code transcript (trimmed):

You: Create three identical worker VMs in zone 1, name them worker-1/2/3, put them behind a new security group that allows SSH from my current IP only, and email me the IPs when they’re running.

Claude: I’ll use the excloud-cli skill. First, let me check what’s already in place — your default subnet, your active SSH key, your public IP — then create the security group and the three VMs. I’ll confirm before creating anything billable.

(runs exc compute subnet list, exc compute key list, exc compute localip)

Claude: Plan:

1. Create security group worker-sg with one ingress rule: TCPv4, port 22, from 203.0.113.42/32.
2. Create three VMs worker-1, worker-2, worker-3 (m1a.large, Ubuntu 24.04, subnet 1, key main, ephemeral public IPv4).
3. Email you the resulting IPs from the mail skill.

Approve?

…and so on. Every CLI command is shown verbatim, so you can stop the agent and run the same commands yourself if you’d rather.

Authentication#

The skills assume you’ve already logged the CLI in:

exc login

They’ll detect a missing token and prompt you to authenticate rather than failing mid-workflow.

For CI environments — where the agent runs unattended — set EXCLOUD_API_KEY and EXCLOUD_ORG_ID in the environment. The CLI (and therefore the skill) picks them up automatically.

Disabling or scoping the skill#

Per-project: install without --global so a skill only fires inside that project’s directory.

Per-action: agents respect the safety guards baked into each skill — excloud-cli will not run destructive commands without confirmation. If you want stricter guardrails, copy the skill locally and edit its policy block.

Source#

The bundle lives at git.excloud.dev/excloud-dev/excloud-skills. Issues, PRs, and skill suggestions welcome.

Related#

• CLI Installation — installs exc plus the skills bundle in one command.
• CLI Overview — what the agent will actually call.
• Terraform Provider — for everything you’d rather declare in code than orchestrate with an agent.