Monitors for your work stack

CoIsland's monitors for your work stack are checks that run from your Mac on a schedule. Each one asks a single question of one connector, in that tool's own language, such as SQL for Snowflake, a GitHub search or JQL for Jira. When the answer holds something new, the island opens, a sound plays and an alert is written, ready to acknowledge, resolve or hand to an AI agent.

A monitor runs on one of the connectors listed in Connectors, every 15 minutes unless you choose otherwise. The check goes straight from your Mac to the tool's own API: no CoIsland server, no CoIsland account.

Connector	Monitor kinds	You write
Gmail	Unread mail, Mail from people, Gmail search, Unread count	A Gmail search, as in Gmail's search box
Calendar	Meetings, Invitations and changes	The calendars to read, the minutes of warning and what to leave out
Dock badges	New unread, Badge over a threshold	The apps to watch, and for a threshold, the count to pass
Apple Mail	Unread mail, Mail from people	The mailboxes to read, and for mail from people, the addresses or domains
GitHub	Issues and pull requests, Review requests, Failed workflow runs, New releases, Security alerts	A GitHub search, or filters such as repo: and branch:
Vercel	Deployments, Stuck builds, Domain problems	A team and projects from your account, then the states or branches that count
GitLab	Review requests, Merge requests and issues, Failed pipelines, New releases, To-do items	Whose merge requests or issues, which projects, pipelines, releases or to-dos
Cloudflare	Failed Pages deployments, Worker errors, Zone problems	An account, then Pages projects, Workers and an error count, or every zone
Render	Failed deploys, Suspended services	Services from your workspaces, or a workspace to watch for suspended services
Docker	Containers down, Containers over a limit	Containers by name:, project: or image:, and for a limit, memory: or cpu:
Local dev	Port conflicts, Ports in use, Local server down, Heavy processes	Ports such as port:3000-3999, a program, a local address, or a memory or CPU limit
Jira	Issues, Status changes, Comments, SLAs, Sprint scope	JQL, as in Jira's advanced search
Linear	Issues, Status changes, New in triage, SLA at risk, Inbox, Project health	A Linear issue filter, or fields such as teams, states and labels
Confluence	Search (CQL), Mentions, Page updates, Comments, Tasks	CQL, or fields such as spaces and a look-back
Asana	Assigned to me, Due and overdue, New tasks in a project, Comments on my tasks	Your tasks, a project or comments, with the workspace and projects picked by name
ClickUp	Assigned to me, Due and overdue, New tasks in a list	Your tasks by status or due date, or a list, with Spaces and lists picked by name
Todoist	Due and overdue, Filter, New tasks in a project	Any Todoist filter, your due tasks, or a project picked by name
Notion	Database rows, Edited pages, Comments	A database picked by name and Notion's filter JSON, or the pages whose comments to watch
Shortcut	Stories, Entered a state, Iteration ending	A Shortcut search, as in its search bar, or a workflow state picked by name
Airtable	Records, Changed records	A base, table and view picked by name, and an Airtable formula
Reminders	Due soon and overdue, New reminders	The lists to read, the minutes of warning and a priority
Sentry	Issue search, Regressed & escalating, Assigned to me, Crons, uptime & metric, New releases	A Sentry search, with projects and environments
PagerDuty	Assigned to me, Incidents, Status changes, Acknowledged too long, On call now, Shift starting soon	Services or teams, picked by name, or your own on-call shifts
Datadog	Monitors alerting, Incidents, Metric threshold	States and tags, severities, or a metric monitor's query and a threshold
New Relic	NRQL query, Open issues	NRQL, or the priorities of the alert issues to watch
Langfuse	LLM errors, Cost over a limit, Slow LLM calls, Eval score below a limit, Negative user feedback	A lookback and the models or environments that count, a limit in dollars or seconds over a window, or a score and its limit or negative values
Snowflake	Custom SQL, Validation	SQL starting with SELECT, WITH or SHOW, or a table, a column and a condition
Databricks	Custom SQL, Validation, Failed job runs, Failed pipeline updates, Triggered SQL alerts	SQL on a SQL warehouse, a validation rule, or the jobs, pipelines and SQL alerts to watch
dbt	Failed job runs, Long runs, Source freshness	The jobs, results and lookback to watch, or a time a run must not exceed
PostHog	HogQL query, Insight threshold, New exceptions	HogQL, or an insight and a number, or how far back to look for new exceptions
Supabase	Read-only SQL, Project health, Advisor lints	SQL run as the read-only role, or the services and advisors to watch
PostgreSQL	Custom SQL, Validation	SQL, one read-only statement, or a validation rule
MySQL	Custom SQL, Validation	SQL, one read-only statement, or a validation rule
ClickHouse	Custom SQL, Validation, Failed queries, Long-running queries, Stuck mutations, Too many parts, Replication lag, Low disk space	SQL sent read-only, a validation rule, or a threshold on failed queries, long queries, mutations, parts, replicas or disks
HTTP / JSON	JSON value, New JSON items, Web page	An address, then a JSON path such as $.status and a rule, a list and its key, or a phrase a page says
RSS / Atom	New entries, Keywords	A feed's address, and optionally keywords
Status pages	Incidents, Component status, Overall status	A listed service or any status page, then the impact, components or level that count
Shell command	Output lines, JSON rows, Command fails	A command line: the program and its arguments, run on your Mac
Stripe	Failed payments, New disputes, Subscriptions, Failed payouts, Low balance	How far back to look, which subscription changes, or the balance to stay above

Each kind knows what makes a result new. Validation writes the SQL for you from a rule such as "order_amount is greater than 1000". A failed GitHub run counts once per attempt, so a re-run that fails again alerts again. A Jira status change is an issue in a status it was not in at the last check.

Test runs the query once against the real tool and shows what matches now, such as "3 issues match now", with a preview. Nothing is saved, compared or alerted, so a test never moves the monitor's baseline.

Advanced sets the columns that identify a row, a row line for the island where {COLUMN} becomes its value, and for Snowflake the warehouse, database and schema. A check wakes a suspended warehouse that has auto-resume on, and Snowflake bills at least 60 seconds per resume, so a small warehouse with a short auto-suspend keeps checks cheap. Snowflake's query history shows each check under the query tag coisland: and the monitor's name.

Each row gets a key: a SHA-256 fingerprint of the columns that identify it, which you can name. Otherwise the kind picks them (an issue's ID, a run's ID and attempt), and Custom SQL uses the whole row.

1. The first check is a silent baseline: it records what already matches and alerts on nothing.
2. Each later check compares its keys with the previous complete one. Rows with a new key raise one alert together.
3. To compare, only fingerprints are kept, in a private file on your Mac, never row values.
4. A partial answer is never compared: past 10,000 Snowflake rows or 100 GitHub results, the monitor asks you to narrow the query.
5. A changed question starts over: a new query, key, connector or rule means a new baseline. A new name, interval or sound does not.

Two alert rules

New rows appear, the default, alerts on every row that entered since the last check.

Count crosses a threshold compares how many rows match with a number (greater than, at least, less than, at most, equal to, not equal to). It fires once when that becomes true, and again only after it clears. There is no silent first check: a table that is already broken says so at once.

• Interval. Minutes, hours or days in the form; a watch file can also use seconds, down to 30.
• On the clock. Checks land on round multiples of their interval, so monitors that share one check together and a Snowflake warehouse wakes once for them all.
• Sleep. Checks pause while your Mac sleeps or is locked. On wake, a monitor that missed checks runs once.
• Busy tools. When a tool says to slow down, the monitor backs off, from one minute up to 15.
• Sound. Glass by default, any sound in your Mac's sound folders, or None. Monitors that fire together open the island once, with one sound.

Each monitor is either Always in the notch or Only when it alerts. New monitors get Always until four are set that way, so a new one never pushes out one you chose. Change it from the chip by its name.

The island holds as many rings as fit along most of your screen's edge. When they do not all fit, monitors with new rows come first, then those in error, then those set to Always, and the last ring becomes a more ring, +N, that opens the list. A monitor already in the island that alerts keeps its place and turns yellow.

Chip	What it means
In notch	It has a ring in the island right now.
On alert	Only when it alerts, and nothing to show: it joins with new rows or an error.
No room	Always, but the island is full: it joins when a place frees up.

Rows also show a line of matching rows since CoIsland opened, the count now, and a state: OK, +N new, Checking, Waiting or Error. Filters count All, New and Errors.

A monitor in error

A monitor that cannot run says why, such as a SQL compilation error or a missing key column. Its ring turns red, it joins the island whatever its mode, and its page opens on the message. The error clears at the next good check, or once the file is fixed, and the monitor keeps what it had seen.

The header holds Check now, and Mark as seen while there is something new. Below it:

• Four figures: rows matching now, new rows not seen yet, the last check and a countdown to the next.
• A chart of every check since CoIsland opened, with a yellow point where rows entered.
• Its alerts, the newest five, each a click away from its own page.
• Its rows, new ones marked NEW, with a filter, Copy CSV and Export.

Dashboards add up alerts, response times and check success across every monitor.

Settings are -- key: value comment lines on top, the query below. Edit a file in any editor and the monitor follows within a second; the form writes the same header you would. This one alerts on every Snowflake task that failed in the last hour:

-- name: Failed tasks
-- connector: acme
-- every: 5m
-- alert: new-rows
-- key: QUERY_ID
-- title: {NAME} failed: {ERROR_MESSAGE}
SELECT query_id, name, error_message, scheduled_time
FROM TABLE(INFORMATION_SCHEMA.TASK_HISTORY(
  SCHEDULED_TIME_RANGE_START => DATEADD('hour', -1, CURRENT_TIMESTAMP()),
  RESULT_LIMIT => 1000))
WHERE state = 'FAILED'

Other connectors add a kind line. Without connector, the tool's default connector runs it:

-- name: Reviews waiting on me
-- kind: github.review-requests
-- every: 5m

user-review-requested:@me is:open draft:false

Every key is in Watch files. New to CoIsland? Start with Getting started.