Lighthouse
Lighthouse v1.0.1
Watch your business metrics
From the developer
Declared capabilities
- Read
- Write
Declared limitations
- Unrelated request
- Data outside the signed-in organization
- Delete without a clear target
Not specified in the available declarations: pricing & account requirements.
Onboarding example
Use Lighthouse to show which of my metrics are alerting right now.
Sources & dates
Developer declarations; not independently tested. Listing last checked Nov 18, 2026. Package observed Nov 18, 2026.
Description
From the marketplace listing
Lighthouse watches the business metrics that live in your data (Snowflake, BigQuery, Postgres, SQL Server, Mixpanel and more) and alerts your team in Slack or email when something changes. With Lighthouse in ChatGPT you can build a metric from a plain-language goal, test and backtest it before it goes live, investigate an alert and acknowledge or resolve it, fix a failing metric, tune a noisy threshold, explore your tables and data sets, and connect new sources through a secure Lighthouse link so credentials never pass through the chat. Admins can also manage members, invitations and business groups. You sign in with your Lighthouse account and pick one organization. ChatGPT can only do what your own Lighthouse role allows. Admins can pause agent access for the whole organization, and every user can disconnect from Settings, API & Agents in Lighthouse. Requires a Lighthouse account.
Language: English · Automatically detected from descriptions.
Changes
No changes recorded yet. Follow this plugin to keep track of future changes.
First seen Nov 18, 2026. Last seen Nov 18, 2026.
Similar plugins
Similar plugins could not be prepared right now. Please try again in a moment.
Browse plugins →Files & skills
File archives
Skill instructions
admin-members-and-groups4.09 KB
--- name: admin-members-and-groups description: Help a Lighthouse organization admin manage members, invitations, roles, business groups and organization settings, confirming every change before making it. Use when the user asks who is in their Lighthouse organization, or to invite, remove or change the role of a member, resend or revoke an invitation, manage or delete business groups, or change organization settings. --- # Manage members, groups and settings These tools are for organization admins only, and every call checks it. Each write changes what real people can do, or sends them an email, so read first, show the exact change, and wait for a yes before each write. If a tool named here is missing, the user's key or role does not allow it. Say which one and that it can be changed in Lighthouse under Settings, API & Agents. ## Read - Members, roles and pending invitations: `list_organization_members` (set includePendingInvitations to see invitations). - Pending invitations with their ids: `list_invitations`. - Find a person's user id by name or email: `search_organization_users`. - Business groups and their members: `list_business_groups`. Groups decide who sees the metrics and data sources shared with them. - Organization settings (time zone, error notifications, metric suggestions): `get_organization_settings`. ## Write (one confirmation per change) - **Invite**: `invite_user` with the email and role (admin or member). It sends a real email that cannot be recalled. Check the address is not already a member or invited, confirm the exact address and role, and call it once. - **Change a role**: `change_member_role`. Takes effect at once. It cannot change the user's own role or demote the last admin. - **Remove a member**: `remove_member`. They lose access at once and can only come back by a new invitation; their metrics stay. Confirm the person by name and email first. - **Resend an invitation**: `resend_invitation`. Clerk has no resend, so the old invitation is revoked and a new real email goes to the same address with the same role. It counts toward the invite limit. Confirm the address first. - **Revoke an invitation**: `revoke_invitation`. The link in the email stops working at once. - **Create a group**: `create_business_group` with members who already belong to the organization (ids from `search_organization_users`). It does not invite anyone. - **Change a group**: `update_business_group` with only what changes. Removing a member can take away their access to what is shared with the group; a group keeps at least one member. - **Delete a group**: `delete_business_group`. It cannot be undone through MCP. It is refused while metrics or data sources are still shared with the group; remove the group from their access first. - **Delete a data set or data source**: `delete_data_set`, `delete_data_source`. They cannot be undone through MCP, and a deleted data source needs its credentials entered again to come back. Each is refused while something still uses it, so clean up in order: metrics (`delete_metric`), then data sets, then the data source. Data sets Lighthouse generated for a data source go with it and cannot be deleted on their own. - **Settings**: `update_organization_settings` for the time zone, error notifications and metric suggestions. The AI model, error destination, billing and the MCP switch are changed in the Lighthouse web app. For a list of changes, confirm the whole list, then make them one at a time and stop at the first error. ## Safety rules - Never write without the admin's yes for that change. A request like "clean up the members" is not a yes to remove anyone in particular; list who you would remove and ask. - Member lists hold names and email addresses. Show them only to the admin who asked; do not copy them anywhere else. - A role change can take up to a minute to apply to some tools. If a call says the user is not an admin right after a change, wait and retry once. ## Report back - Eac
Preview of the first 4,000 characters. Read full instructions →
build-metric4.29 KB
---
name: build-metric
description: Build a new Lighthouse metric (a scheduled check on the user's data that alerts in Slack or email when a threshold is met) from a plain-language goal, test it, backtest it and take it live. Use when the user asks to monitor, track, watch or alert on something in their data with Lighthouse.
---
# Build a metric
A metric runs a query on a data source on a schedule, compares the result with a threshold and notifies people when it is met. Build it as a DRAFT, prove it works, then go live only when the user says so.
If a tool named here is missing, the user's key or role does not allow it. Say which one and that it can be changed in Lighthouse under Settings, API & Agents.
## Steps
1. **Pin down the goal.** What number, per what (segments such as country or customer), how often, what counts as bad, and who should hear about it. Ask only for what you cannot find.
2. **Find the data.**
- `list_data_sources` and `list_data_sets` for what exists; `get_data_set` for column ids and which columns can be a metric, filter, segment or date.
- Prefer a CUSTOM metric on a data set: it gets trend and threshold options and can be backtested. Use RAW_SQL only when no data set fits and one should not be made.
- If no data set fits, follow the explore-data-and-model-datasets skill (`list_data_source_tables`, `list_data_source_columns`, `query_data_source`, `create_data_set`).
3. **Start from an example.** Call `search_metric_configuration_examples` with words from the goal (for example "count", "segments", "trend max", "calculation window", "slack"), then peek at up to 5 example ids for the full JSON. Copy the closest shape; do not invent fields.
4. **Assemble the configuration** with real ids only:
- data source and data set ids from step 2;
- a schedule (every metric needs one);
- notifications: `list_communication_sources`, then `get_communication_source` for Slack channel or user ids. Email recipients must be members of the organization;
- KPI tags if the user wants them (`list_kpis`);
- who may edit and view it. Ask the user; do not assume. For editing: everyone (`editAccessLevel` `EVERYONE`, "Everyone" in the web app) or only them (`ADMINS`, shown as "Owner"). For viewing, the same choice with `viewAccessLevel`. For specific people or groups, use `SELECTED_USERS` with editorUserIds, viewerUserIds, editorBusinessGroupIds or viewerBusinessGroupIds (ids from `search_users_and_groups`, which anyone who can write metrics may use). Pass the choice in the configuration.
5. **Test.** Call `test_metric` with the full configuration. Fix every error and repeat until valid is true. Show the user the preview rows and check they look like the number they meant.
6. **Save as a draft.** `create_metric` with status DRAFT. A draft never runs on its schedule and sends nothing.
7. **Backtest.** For a CUSTOM metric with a calculation function, call `preview_metric_history` with the new metric id. Report how often the threshold would have been met per day. If it would fire too often or never, adjust the threshold with `test_metric` and `preview_metric_history` (pass metricId and only the changed fields), then save with `update_metric`. See the tune-alert-noise skill for more.
8. **Go live, with consent.** Summarize what will run, who will be notified and who can edit and view it, and ask. Only after a clear yes, set status LIVE with `update_metric`. LIVE counts toward the plan's live metric limit.
## Safety rules
- Ask before `run_metric`: a run can send the metric's real notifications and actions, like a scheduled run. `test_metric` and `preview_metric_history` never notify anyone.
- Never call `send_test_notification` unless the user asked for a test message. It sends a real message to real people; call it once.
- `test_metric` and `preview_metric_history` query the customer's warehouse, which can cost money and is rate limited. Do not loop on them; change something between calls.
- Do noPreview of the first 4,000 characters. Read full instructions →
connect-sources3.11 KB
---
name: connect-sources
description: Connect a data source (Postgres, Snowflake, BigQuery, SQL Server, HTTP connector, Mixpanel), a Slack or email notification destination, or an integration action source to Lighthouse, or reconnect one with new credentials, through a secure Lighthouse link so credentials never pass through the chat. Use when the user wants to connect, add, reconnect or fix the credentials of a source in Lighthouse.
---
# Connect sources
Credentials never go through this conversation. The begin tools return a Lighthouse link; the user opens it while signed in, enters the details there and saves. You then check how it ended.
If a tool named here is missing, the user's key or role does not allow it. Say which one and that it can be changed in Lighthouse under Settings, API & Agents.
## Steps
1. **Check what exists.** `list_data_sources` or `list_communication_sources`, so you do not create a duplicate. To reconnect, find the id with `get_data_source`.
2. **Start the setup.**
- Data source: `begin_data_source_setup` with the dataSourceType (POSTGRES, SNOWFLAKE, BIGQUERY, MSSQL, HTTP_CONNECTOR, MIXPANEL) to create, or the dataSourceId to replace credentials or connection settings.
- Slack or email notifications: `begin_communication_source_setup` with type SLACK (the user approves Lighthouse on Slack's own page) or EMAIL (an organization can have only one).
- A place Lighthouse sends actions to, such as an HTTP endpoint: `begin_integration_action_source_setup`, with the integrationActionSourceId to reconnect one.
3. **Hand over the link.** Give the user the url and say: open it while signed in to Lighthouse in the same organization, fill in the details, save. It works for 20 minutes and only for them. For a warehouse, suggest a read-only database user or role.
4. **Check the result** when the user says they are done: `get_setup_status` with the setupId.
- completed: entityId is the new or updated source.
- pending or in_progress: they have not saved yet, or the save is running. Wait and check again.
- expired: start again for a new link.
- unknown: the save may or may not have happened. Check `list_data_sources` or `list_communication_sources` (or Lighthouse itself, for an integration action source) before starting another setup.
5. **Tidy up.** Rename or describe a source with `update_data_source` or `update_integration_action_source`. These never change credentials or connection settings.
6. **Next step.** For a new data source, offer to explore it (explore-data-and-model-datasets skill).
## Safety rules
- Never ask for a password, key, token or connection string in the chat, and never put one in a tool argument.
- If the user pastes a secret anyway, do not use it. Tell them to use the link instead and to rotate that secret, since it is now in the chat history.
- Calling a begin tool again for the same target returns the same link while it is waiting; do not spam new setups.
## Report back
- What was connected or reconnected, its name and id, from `get_setup_status`.
- Anything left for the user (a link still waiting, an expired link, a Slack approval).
- Suggested next step.
explore-data-and-model-datasets3.18 KB
---
name: explore-data-and-model-datasets
description: Explore the tables and columns in a Lighthouse data source, run narrow read-only queries to understand the data, and create or adjust Lighthouse data sets that metrics are built on. Use when the user asks what data Lighthouse can see, what could be monitored, or to model a table as a data set.
---
# Explore data and model data sets
A data source is a connection to a warehouse or app. A data set is one of its tables, with roles for each column (metric, filter, segment, date). CUSTOM metrics are built on data sets.
If a tool named here is missing, the user's key or role does not allow it. Say which one and that it can be changed in Lighthouse under Settings, API & Agents.
## Steps
1. **See what exists.** `list_data_sources`, `get_data_source`, `list_data_sets` and `get_data_set`. Reuse a data set that already covers the table.
2. **Browse the schema (free).** `list_data_source_tables` and `list_data_source_columns` run no warehouse query. Search by name; page through large schemas.
3. **Look at real values (costs a query).**
- For an existing data set, `peek_data_set` with its column ids and a date range.
- Otherwise `query_data_source` with one narrow SELECT: only the columns you need, a date filter, aggregates or a small LIMIT. Copy table and column names from step 2.
- Both work on warehouse data sources (Postgres, Snowflake, BigQuery) only.
- Each query runs on the customer's warehouse, can cost money, and shares a per-minute limit (10 by default) with `peek_data_set` and `preview_metric_history`. Plan queries; do not fire them in a loop.
4. **Propose a data set.** Name the table, the default date column (must be a date or timestamp), which number columns can be metrics, and which text columns are segments. Show this to the user.
5. **Create it, with consent.** `create_data_set` from one table. Leave out columns to take every column with default roles, or list the columns you want. Data sets defined by custom SQL are not supported.
6. **Adjust it.** `get_data_set` first, then `update_data_set` with only what changes.
7. **Next step.** Offer to build a metric on it (build-metric skill).
## Safety rules
- Query results are the customer's data. Show only what answers the question; never copy rows anywhere the user did not ask for.
- `query_data_source` only accepts one read-only SELECT or WITH. If it refuses a query, fix the SQL rather than working around the check.
- Before `update_data_set` removes a column from metrics or takes a role away, warn the user: metrics built on the data set change with it and can start failing. Ask for a yes.
- Recommend a read-only database user for the data source if the user is unsure what it can reach.
- Deleting a data set or data source is for organization admins (admin-members-and-groups skill). Never delete one to start over; adjust it with `update_data_set`.
## Report back
- What the data source contains that matters for the user's goal (tables, key columns, grain, date range).
- Findings from queries, as short numbers or a few rows, with the SQL you ran.
- Data sets created or changed: name, id, date column, metric and segment columns.
- Metrics this suggests, if the user wants them.
fix-failing-metric3.54 KB
---
name: fix-failing-metric
description: Diagnose and fix a Lighthouse metric whose runs fail or that Lighthouse marked INVALID, using its execution errors and change history, then test the fix and resume it with the user's agreement. Use when a metric is failing, erroring, INVALID, stopped running, or shows no new results.
---
# Fix a failing metric
Every failed run adds to a metric's failure count; when a LIVE metric reaches the limit (3 by default) Lighthouse sets it INVALID and stops scheduling it. Find the cause, fix it, prove the fix without notifying anyone, then resume.
If a tool named here is missing, the user's key or role does not allow it. Say which one and that it can be changed in Lighthouse under Settings, API & Agents.
## Steps
1. **Find the metric.** `list_metrics` by name or sequential id, then `get_metric_configuration`. Note whose metric it is (see the safety rules).
2. **Read the failures.** `search_metric_executions` for the metric over the recent range, looking at failed runs. Read error and errorHint, and note the first failed run and its metricVersionNumber.
3. **Sort the cause.**
- **Temporary** (errorHint says the source may be rate limiting, a timeout, a 5xx): nothing to fix in the metric. Tell the user and suggest waiting.
- **Credentials or connection** (authentication failed, expired key, unreachable host): the data source needs new credentials. `get_data_source` to confirm which one, then follow the connect-sources skill to reconnect it with `begin_data_source_setup`. Never ask for the credential in chat.
- **Metric changed**: `get_metric_change_history` and compare the versions around the first failure. A field changed just before the failures started is the prime suspect.
- **Data changed** (unknown column, missing table): `get_data_set` for a CUSTOM metric (a column dropped or a role taken away), or `list_data_source_columns` and a small `query_data_source` for RAW_SQL.
4. **Fix it.** Build the smallest change. Call `test_metric` with metricId and only the changed fields until valid is true and the preview looks right.
5. **Save, with consent.** Show the user the change. `update_metric` with the changed fields. For an INVALID metric, setting status LIVE resumes it and resets the failure count; ask before doing that, because it starts scheduled runs and alerts again.
6. **Verify.** The next scheduled run will show the result in `search_metric_executions`. Only if the user wants proof now, and knows a run can send real notifications, call `run_metric` and poll `check_metric_run_status` until resultsAvailable or DONE.
## Safety rules
- **Someone else's metric.** `get_metric_configuration` returns `ownership` with the owner's and creator's names and `isOwnedByCaller`. If the metric belongs to someone other than the user, tell the user whose metric it is and ask before changing it. `update_metric`, `run_metric`, `delete_metric` and `act_on_segment_result` on such a metric change nothing until called again with `confirmedNotOwner: true`; set it only after the user confirms.
- Do not call `run_metric` just to check: `test_metric` proves the query without notifying anyone.
- Do not set LIVE without the user's yes.
- Never delete and recreate a metric to "fix" it; `delete_metric` removes its history for good.
## Report back
- What failed, since when, and the root cause with evidence (error text, the version that changed it).
- The exact change made, and the test result.
- The metric's status now, and whether a manual run was done.
- Anything the user must do (for example re-enter credentials through the link).
investigate-alert3.15 KB
--- name: investigate-alert description: Investigate a Lighthouse alert, find out what fired, whether it is still firing, whether it is real and what likely caused it, then acknowledge, resolve or mark it as a hit when the user agrees. Use when the user mentions a Lighthouse alert, a metric that fired, a Slack alert from Lighthouse, or asks why a metric is alerting. --- # Investigate an alert An alert is a segment result whose value met the metric's threshold in a run. Work out what happened from stored results first; query the warehouse only when the stored data cannot answer. If a tool named here is missing, the user's key or role does not allow it. Say which one and that it can be changed in Lighthouse under Settings, API & Agents. ## Steps 1. **Find the metric.** `list_metrics` by name or sequential id (for example 153). Then `get_metric_configuration` to read the threshold, segments, schedule and notifications. 2. **Is it alerting now?** `get_current_metric_matches` for the segments that match right now. 3. **What fired?** `search_metric_segment_results` for the metric, narrowed by time range, segments or one run (metricRunId), and to results where the threshold was met. Read value, the threshold, segmentValues and resultColumns (baselines and rule thresholds). Set includeRuleResults to see which rule was met. 4. **Is it unusual?** `get_metric_value_history` for the segment over the last 30 days: compare the value with the daily mean and spread in summary.segments, and check runHealth for failed runs. 5. **Did something change?** `get_metric_change_history`. Compare the result's metricVersionNumber with the versions; a recent threshold, filter or SQL change often explains a new alert. 6. **How was it handled before?** `search_metric_activity` shows who acknowledged, resolved or marked results as hits, and how often past alerts were real. 7. **Run health.** If results look missing or stale, `search_metric_executions` for failed runs and their errorHint. Failing runs belong to the fix-failing-metric skill. 8. **Dig into the data only if needed.** `peek_data_set` (CUSTOM) or a narrow `query_data_source` filtered to the segment and the alert window. Both query the customer's warehouse, can cost money and are rate limited. 9. **Act, with consent.** Propose one action and wait for a yes before calling `act_on_segment_result` with the segmentRunId: `ACKNOWLEDGE`, `MARK_AS_RESOLVED` or `MARK_HIT`. To reverse one, use its undo, such as `UNDO_ACKNOWLEDGE`. ## Safety rules - `act_on_segment_result` changes the alert's state for the whole team and edits the Slack message already sent for it. Never call it without the user's yes for that specific result. - Do not call `run_metric` to "refresh" an alert unless the user asks: a run can send notifications again. - Treat query results as the customer's data. Quote only the rows needed to explain the cause. ## Report back - Metric, segment, when it fired, value against threshold, and whether it is still alerting. - Likely cause with the evidence (history, recent changes, data rows), and how sure you are. - What the team did with similar alerts before. - The action you took, or the one you recommend and are waiting on.
tune-alert-noise3.37 KB
--- name: tune-alert-noise description: Make a noisy or silent Lighthouse metric alert at the right rate by measuring its history, backtesting candidate thresholds or filters, and applying the best one with the user's agreement. Use when the user says a metric alerts too often, too rarely, is noisy, has false positives, or asks to tune or adjust a threshold. --- # Tune alert noise Measure first, then backtest options side by side, then change one thing. A good threshold fires on the cases the team treats as real and stays quiet otherwise. If a tool named here is missing, the user's key or role does not allow it. Say which one and that it can be changed in Lighthouse under Settings, API & Agents. ## Steps 1. **Read the metric.** `list_metrics`, then `get_metric_configuration` for the threshold, filters, segments, trend settings and schedule, and whose metric it is (see the safety rules). 2. **Measure the noise.** `get_metric_value_history` over 30 to 90 days. Use summary.segments: threshold-met runs and days, mean, standard deviation and coefficient of variation per segment. Note which segments cause most of the hits. 3. **Check what the team thought.** `search_metric_activity`: many acknowledged or resolved results and few marked as hits means false alarms; hits marked real are alerts to keep. 4. **Check past attempts.** `get_metric_change_history` to see whether the threshold was already moved and what happened. 5. **Look at a few alerts.** `search_metric_segment_results` with the threshold met and includeRuleResults, to see which rule fires and by how much. 6. **Backtest candidates.** For each option (a higher threshold, a trend comparison, a minimum volume filter, excluding a segment), call `preview_metric_history` with metricId and only the changed fields. Use `search_metric_configuration_examples` for the right JSON shape. Each backtest queries the warehouse twice, can cost money and is rate limited, so try two or three options, not ten. Only CUSTOM metrics with a calculation function can be backtested. 7. **Recommend one.** A small table: option, hits in total, days with hits, which past real hits it keeps. 8. **Apply, with consent.** `test_metric` with the chosen change, then `update_metric`. To trial a change without touching the original, `duplicate_metric` makes a DRAFT copy that sends nothing until it is made LIVE. ## Safety rules - **Someone else's metric.** `get_metric_configuration` returns `ownership` with the owner's and creator's names and `isOwnedByCaller`. If the metric belongs to someone other than the user, tell the user whose metric it is and ask before changing it. `update_metric`, `run_metric`, `delete_metric` and `act_on_segment_result` on such a metric change nothing until called again with `confirmedNotOwner: true`; set it only after the user confirms. - Do not change a LIVE metric without the user's yes on the exact change. - `send_test_notification` sends a real message to the metric's channel or recipients. Call it only when the user asks to see the new alert, and only once. - Do not use `run_metric` to test a threshold: it can send real alerts. Use `preview_metric_history` or `test_metric`. ## Report back - Current noise: hits per day and the segments behind them, and how the team treated them. - The options you backtested with their numbers. - The change applied (or recommended), and the expected alert rate. - Any DRAFT copy you created, with its id.
Release notes
First release of the Lighthouse plugin for ChatGPT: browser sign-in to the hosted Lighthouse MCP server, tools to read, test and manage metrics, data sources, alerts and members, and skills for building metrics, investigating alerts, fixing failing metrics, tuning alert noise, exploring data, connecting sources and managing members.
Declared in the saved package. Remote tools may change independently.
Package details
Publisher declarations from the archived package. These are separate from our research and the live service's terms.
- Review scenarios
- 5 positive · 3 negativeScenario counts, not test results.
Declared capabilities
- Read
- Write
Package observed Nov 18, 2026.
Technical details
- First seen
- Nov 18, 2026 · 06:00 UTC
- Last seen
- Nov 18, 2026 · 06:00 UTC
- Collection status
- Collected
plugin_asdk_app_6ada4c9309a08191b66e19ed1069454c
Download plugin data (JSON)Plugin details: freshly prepared. Prepared . Refreshed when plugin data changes or after 6 hours. Your account and Follow status are live.