# Set up CoIsland with your AI

> Let Claude Code, Codex or Cortex Code create your monitors and triage your alerts, with two open-source skills.

Source: https://coisland.app/docs/ai-setup/

## What you need

- CoIsland on this Mac, with at least one connector.
- Claude Code, Codex, Cortex Code or another agent that reads `SKILL.md`.

Three agent skills let the AI you already use set CoIsland up and run it for you. Ask Claude Code, Codex or Cortex Code to "watch failed GitHub Actions runs on acme/api", and it writes the monitor, checks it and tells you the result. Ask "what fired overnight?", and it sums up your open alerts and resolves the ones you call noise. The skills are open source, in the [coisland-community repository](https://github.com/DonTizi/coisland-community) on GitHub.

This is the other way round from [Agents](https://coisland.app/docs/agents/), where CoIsland hands one alert to an agent in a terminal under it. Here you start in your agent, and it works on CoIsland.

## What the skills do

Each skill is a folder with a `SKILL.md` file, the open format Claude Code, Codex and Cortex Code all read. Your agent loads one when your request matches it.

- **`coisland-monitors`** creates, edits, pauses and deletes monitors, and checks them. It lists your connectors by name and tool, writes the monitor's [watch file](https://coisland.app/docs/watch-files/), then runs CoIsland's check on it and reports the result.
- **`coisland-alerts`** sums up your open alerts, new or acknowledged, by monitor, with the rows or items that fired them. It acknowledges, resolves and adds notes.
- **`coisland-support`** reads CoIsland's diagnostic report (`CoIsland --report`, from 0.4.6), tells you what it says, and files a bug in the public issue tracker only after you have read the issue and said yes.

Both work the way the app does:

- A monitor is a plain file in `~/.coisland/watches`, so CoIsland picks up the agent's work at once, as it does your own edits.
- A deleted monitor goes to the Trash.
- The monitors skill refuses SQL that writes, as the app does.
- Alert changes go through the running app, the same way a click on **Resolve** does, so the notch and **Settings › Alerts** update at once. With CoIsland closed, the change is written to its files, and the app shows it when it next opens.

## Install the skills

### Paste a prompt into your AI

The fastest way: paste this into Claude Code, Codex, Cortex Code or another agent. It installs the skills, checks them, lists your connectors, then asks what you want to watch and creates your first monitors with you:

```
Install the CoIsland skills from https://github.com/DonTizi/coisland-community, following its README: run curl -fsSL https://raw.githubusercontent.com/DonTizi/coisland-community/main/install.sh | sh, or copy its skills folder into your own skills folder.
Then check that coisland-monitors, coisland-alerts and coisland-support are installed, and tell me.
Skills load when a session starts, so read the installed coisland-monitors SKILL.md and its reference.md now, and follow them for the rest of this session.
Then list my CoIsland connectors with /Applications/CoIsland.app/Contents/MacOS/CoIsland --connectors.
Then ask me what I want to watch, and create my first monitors as that skill says, checking each one.
Never ask for or handle a token. If a connector is missing, open its form in CoIsland with open "coisland://connectors/add?provider=<id>" and wait for me to add it.
If you cannot run commands, tell me the commands to run.
```

The agent still asks before each command, in its own mode. An agent that cannot run commands tells you which ones to run.

### Or from the terminal

One line installs both skills into every agent it finds on your Mac, and prints what it did:

```
curl -fsSL https://raw.githubusercontent.com/DonTizi/coisland-community/main/install.sh | sh
```

Then start a new session in your agent: agents read their skills when a session starts. To install for one agent only, or by hand, use its section below.

### Claude Code

Claude Code reads skills from `~/.claude/skills`. To copy them there by hand:

```
git clone https://github.com/DonTizi/coisland-community.git
mkdir -p ~/.claude/skills
cp -R coisland-community/skills/* ~/.claude/skills/
```

In a new session, type `/` and the two skills are listed among the commands.

### Codex

Codex reads skills from `~/.agents/skills`, the folder its documentation gives. `~/.codex/skills` works too.

```
git clone https://github.com/DonTizi/coisland-community.git
mkdir -p ~/.agents/skills
cp -R coisland-community/skills/* ~/.agents/skills/
```

In a new session, `/skills` lists them.

### Cortex Code

Cortex Code installs them from the repository itself:

```
cortex skill add https://github.com/DonTizi/coisland-community.git
```

Or copy them into `~/.snowflake/cortex/skills`:

```
git clone https://github.com/DonTizi/coisland-community.git
mkdir -p ~/.snowflake/cortex/skills
cp -R coisland-community/skills/* ~/.snowflake/cortex/skills/
```

Cortex Code also reads `~/.claude/skills`. If Claude Code already has the skills, Cortex Code has them too: do not add a second copy, or it loads both. `cortex skill list` shows what it found.

### Any other agent that reads SKILL.md

Download the zip from the [latest release](https://github.com/DonTizi/coisland-community/releases/latest) and unzip it into your agent's skills folder: one folder per skill, each with its `SKILL.md`.

## What to ask your AI

Ask in your own words. For example:

- "Watch failed GitHub Actions runs on acme/api"
- "Alert me when orders over $10,000 land in Snowflake"
- "What fired overnight? Resolve the noise and note why"
- "Pause my Jira SLA monitor for the weekend"

The agent reads which connectors you have, so name the tool or the connector when you have several. Its own mode still asks before each command it runs, as in any session.

## What your AI can and cannot do

The skills give an agent CoIsland's own local surface, and nothing more:

- **It never handles a token.** When a monitor needs a connector you do not have, the skill opens CoIsland's add-connector form for that tool. You type the token there, and it goes to your login Keychain, as always. The agent reads connector names and tools, never secrets.
- **It works only on this Mac,** through CoIsland's command line and your watches folder. There is no CoIsland server, account or remote control.
- **It never edits the app's files directly.** Alerts change through the app, with the same code a click uses.
- **It cannot make a monitor write.** SQL that changes data is refused.

What the agent reads, such as an alert's rows or an issue's comments, goes to its own service under your account, as in any session: Claude Code to Anthropic, Codex to OpenAI, Cortex Code to Snowflake. Pick the agent your company allows for that data.

## Update and uninstall the skills

Each skill names the CoIsland version it describes. When you update CoIsland, update the skills too: run the install line again, or copy them again by hand. With Cortex Code, `cortex skill check` says when the repository has a newer version.

To remove them, delete their two folders from your agent's skills folder. For Claude Code:

```
rm -rf ~/.claude/skills/coisland-monitors ~/.claude/skills/coisland-alerts ~/.claude/skills/coisland-support
```

For Codex, the same with `~/.agents/skills`. For Cortex Code, `cortex skill remove coisland-monitors`, `cortex skill remove coisland-alerts` and `cortex skill remove coisland-support`. Your monitors and alerts stay: they belong to CoIsland, not to the skills.

## Troubleshooting

- **The agent does not pick the skill: name it.** Say "Use the coisland-monitors skill to…", or call it directly: `/coisland-monitors` in Claude Code, `$coisland-monitors` in Codex. Agents choose a skill from its description, and not every phrasing matches.
- **The agent does not know the skills.** Start a new session after installing, and check the folder: each skill must be its own folder, such as `~/.claude/skills/coisland-monitors/SKILL.md`, not one level deeper.
- **Cortex Code needs a Snowflake connection.** Cortex Code runs on a connection from `~/.snowflake/connections.toml`, even for a GitHub monitor. Run `cortex` once to pick one, or add one with `snow connection add`. [Agents](https://coisland.app/docs/agents/) covers its setup.
- **The check fails.** The agent reports CoIsland's own message about the file. [Watch files](https://coisland.app/docs/watch-files/) lists every key and kind, and the monitor's page in **Settings › Monitors** shows the error of its last check.
- **The agent asks before running CoIsland.** That is its mode doing its job: approve the command, or allow it for the session.

Something else? [Open an issue on GitHub](https://github.com/DonTizi/coisland-community/issues/new/choose).

## Questions about CoIsland and your AI

### Can my AI see my tokens?

No. Tokens stay in your login Keychain, and the skills only read connector names and tools. A missing connector opens CoIsland's form, where you type the token yourself.

### Does CoIsland need to be open?

No. Monitors are files, which CoIsland reads when it opens. An alert change made while it is closed goes to its files, and the app shows it when it next opens.

### Do I need the skills to use agents in CoIsland?

No. CoIsland starts Claude Code, Codex or Cortex Code on an alert without them: see [Agents](https://coisland.app/docs/agents/). The skills are for starting from your agent instead.