Alerts from your work stack, in one place on your Mac

Every alert starts with a monitor: one question asked of one tool, in its own language, every 15 minutes by default (from 30 seconds to 30 days). Each monitor has one rule:

• New rows or items appear: something matches that did not match at the previous check.
• The count crosses a threshold: greater than, at least, less than, at most, equal to or not equal to a number you choose.

When a check makes the rule true, CoIsland records one alert with what it saw. Whichever connector ran the check, the alert lands in the same island and the same list, titled in the monitor's own words:

Connector	Monitor and its query	The alert it raises
Snowflake	Custom SQL: orders over $10,000 in sales.orders	3 new rows in Big orders
GitHub	Issues and pull requests: repo:acme/api is:issue is:open label:bug	2 new items in API bugs
Jira	SLAs: project = HELP AND "Time to resolution" = breached()	1 new request in Breached SLAs
Snowflake	Failed tasks, with a count rule: count > 0	Failed tasks: row count greater than 0

The first check of a new-rows monitor raises nothing: it records what already matches as its baseline. A count rule has no silent first check, so a table that already breaks it tells you at once. A check that fails shows as an error on the monitor, not as an alert.

When a monitor fires, the island opens for five seconds on its card and plays its sound: Glass unless you pick another of your Mac's sounds, or none. Monitors that fire together open the island once, with one sound. With the island hidden, the sound still plays.

• The ring shows +N and pulses yellow, and the card lists the new rows or items first.
• A click on the ring opens the monitor's newest new alert.
• Press and hold the ring to mark its rows seen: its new alerts become acknowledged.
• The card ends with Claude Code, Codex and Cortex Code, to start one on the newest new alert.

Your first four monitors stay in the island; later ones join it only when they alert. Each row of Monitors changes that. The home page shows the cards.

Three facts lead the page: Expected, what the monitor was waiting for ("No new matching rows"); Observed, what it saw ("1 new row, 5 matching at that check"); and Detected, when.

• The rows that fired it, as they read when it fired, up to 200. Copy CSV puts them on your clipboard.
• The items, for GitHub and Jira: each issue, pull request or comment as a card, then as the tool has it now, with Open on GitHub or Open in Jira.
• Details: the monitor, linked, its status, rule, connector and, for SQL, tables.
• Activity: when it was raised, every status change, and your notes, each dated.

Delete the monitor and its alerts stay, marked deleted.

Acknowledge, Resolve and Reopen sit at the top of the alert page and in the list's right-click menu. Nothing changes status on its own: CoIsland never resolves an alert because its rows left the result. Three shortcuts acknowledge in bulk; none resolves:

Where	What it does
Mark as seen, on a monitor's page	That monitor's new alerts become acknowledged, and its ring stops counting.
Press and hold a ring, in the island	The same, without opening Settings.
Acknowledge All, in the menu bar menu and in General › Checks	Every monitor's new alerts become acknowledged.

Each change is dated in the activity, so the dashboards know how long alerts wait.

Every alert, whatever its connector, lands in Settings › Alerts. A row shows the monitor's ring in the status color, the title, the first row or item that fired it and the time.

• Filters: All, New, Acknowledged and Resolved, each with its count.
• Search: titles, monitor names and the rows or items that fired them.
• The sidebar counts new alerts, with a yellow dot.

CoIsland keeps the 500 newest alerts in one file on your Mac, ~/Library/Application Support/CoIsland/alerts.json, that only your user account can read. There is no CoIsland server to send them to.

A new-rows monitor raises one alert per check that finds something new: ten failed runs in one check are one alert with ten rows. A row that stays in the result never alerts twice. A count rule fires once when it starts matching, and again only after it has cleared.

This Snowflake watch file alerts on each failed task run once, keyed by task and scheduled time:

-- name: Failed tasks
-- database: ANALYTICS
-- every: 5m
-- key: NAME, SCHEDULED_TIME
-- alert: new-rows
-- sound: Basso
SELECT name, scheduled_time, error_message
FROM TABLE(INFORMATION_SCHEMA.TASK_HISTORY(
  SCHEDULED_TIME_RANGE_START => DATEADD('hour', -1, CURRENT_TIMESTAMP())))
WHERE state = 'FAILED'

With -- alert: count > 0 instead, it alerts once when failures appear, then stays quiet until a check finds none in the last hour. To hear less, give a SQL monitor a key (without one, a row whose columns change reads as new), narrow its query, or set its sound to none.

On the Dashboards, Actionability is the share of alerts someone acknowledged or resolved: a falling share is alert fatigue. Noisiest monitors names the ones raising the most.

At the foot of every alert page, Work on this with an agent offers Claude Code, Codex and Cortex Code. A click starts that agent's own CLI, with your account, in a terminal under the alert.

Its first message is the whole alert: the facts, the monitor's query, every row and item kept, and the activity with your notes. It asks the agent to find the cause, fix what is safe and can be undone, and ask you before anything destructive. Claude Code and Codex start in their auto modes, Cortex Code in its default Confirm mode. The agent adds notes to the alert but never acknowledges or resolves it: you do.

Each session is a tab along the bar and a row in Settings › Agents, to resume, stop or archive. See Agents, then the agents docs to set them up, or start with Getting started.