Plausible

Discover sites and configured goals, then analyze website traffic, pages, acquisition, UTM campaigns and measured goal conversions through the existing workspace app. One bundled skill, plausible-analysis, covers these related workflows and returns results in the user's language. Install the whole plugin so its query and output references remain available. No site is hardcoded.

Usage and output

Ask a website analytics question and name the site when it is not already known. The skill discovers the site's reporting zone and local date with get_sites, checks goal configuration with get_goals, and queries interval KPIs with get_aggregate. It uses paginated breakdowns when needed and separates observations from hypotheses and next checks. Discovery uses the existing caller access; unavailable metadata remains unknown.

The skill and its self-contained query and output references are written in Dutch; results follow the language of the user's request. The seven read tools are:

Tool	Purpose
get_sites	Authorized sites, reporting time zones and site-local today; cursor pagination.
get_goals	Configured goals, including goals with no activity; cursor pagination.
get_aggregate	Interval-wide totals, rates and averages.
get_timeseries	Daily, weekly or monthly trends.
get_breakdown	Ranked, paginated results for one dimension.
get_conversions	Goal activity and conversion rates, optionally per page, with pagination.
compare_periods	Two period aggregates, deltas and per-period metadata and coverage.

• Chat report: concise interpretation and Markdown KPI or breakdown tables (default).
• Graphs and tables: inline charts when the host can render them, with underlying values.
• PDF report: an actual attachment when the host offers file generation or PDF export.

Charts and PDFs depend on host tools; this package supplies reporting instructions, not a rendering engine. Unavailable modes produce an explicit limitation and a chat fallback. There is no required dependency on another analytics, charting or PDF plugin.

Connection and permissions

.app.json maps the required plausible app to the supplied workspace ID asdk_app_6ad0a3e9dc288191a3e84ecfd5b23683. Administrators enable the app for intended roles separately; members authenticate through the host. This package neither deploys the MCP server nor contains credentials or changes workspace policies.

The analytics tools are read-only. Write is also declared because the app exposes send_feedback, which messages maintainers. The skill requires explicit user authorization for that message; routine analysis does not send feedback.

Compatibility and evidence

ChatGPT Business is the intended target. Installation, app mapping, activation, resource loading, live query behavior, charts and PDF export are unverified in that workspace. The app is unavailable in this development chat, so no live analytics queries were run. Both catalogs register the package for discovery; neither establishes runtime support.

The Claude overlay at .claude-plugin/plugin.json connects the plausible HTTP MCP server to https://plausible-mcp.onderwijsin.nl/internal, using the same bundled skills. Authenticate through the host's OAuth flow with an authorized account. The server's OAuth configuration must allow the host's callback; the plugin contains no tokens or client secrets. This endpoint is separate from the OpenAI app mapping, so ChatGPT continues to use .app.json without a second bundled connection.

Claude installation, OAuth, tool activation and representative tasks remain unverified. Codex requires access to the registered app or equivalent authenticated tools. The OpenAI presentation uses assets/icon.png for both the logo and composer icon.

Integration references:

• OpenAI packaging: portable root manifest and extensions.com.openai.apps pointing to .app.json. The supplied asdk_app_... ID is preserved, matching the existing package pattern. Current docs illustrate plugin_asdk_app...; verify the exact mapping during workspace installation instead of guessing a replacement.
• Plausible MCP: tool schemas, filtering and authentication.
• Claude manifest reference: inline mcpServers in .claude-plugin/plugin.json and shared root skills/ discovery.
• Plausible Stats API: the full API contract; this wrapper exposes a subset.
• Metric definitions: measurement terminology.

All five statistics tools accept include_imports; native-only remains the default. Explicitly requested imported history is checked against returned coverage, not assumed from the request. Structured results preserve metadata and resolved queries; comparisons keep them separately for each period. Breakdowns and conversions expose offsets, next offsets, totals, completeness and warnings. A terminal page alone does not prove a complete inventory, and pagination is not a snapshot.

A trailing page wildcard selects a literal prefix: /traineeship* includes /traineeship/apply and excludes /archive/traineeship. This narrows the previous substring scope. Intentional substring queries must use an event:page contains property filter. Deployments without upstream matches support return an error rather than silently widening the page selection.

Discovery depends on the Plausible version and caller permissions. Unsupported or denied discovery is unknown configuration, not evidence of absent sites or goals. Remaining query limits include one breakdown dimension, no arbitrary ordering, no timestamp bounds or time-zone override, no behavioral session filters and no time-label inclusion option. Breakdown pages contain at most 1000 rows; conversion pages at most 10000. The self-contained query guide describes parameters, defaults, pagination, filters, coverage and examples.

Conversions require configured, measurable goals. The skill cannot infer visitor journeys, ad costs or ROAS from traffic, and never treats cross-site visitor sums as deduplicated people.

Manual acceptance matrix

All cases are pending runtime testing. Install in the intended host, confirm app access and automatic activation, and check both bundled references load. Capture actual query parameters and results for each case using authorized test-site data. Repeat for every host for which support will be claimed.

Request	Acceptance criterion
“How did onderwijsin.nl perform in the last 30 days?”	Site zone/local date from get_sites; five KPIs from get_aggregate; exact complete days and interpretation.
“Which pages performed best last month?”	Previous calendar month; ranking metric first in query; visitor/pageview distinction and top-N scope.
“Where did traffic to our traineeship pages come from?”	Verified page set; source/channel and relevant UTM filters retained; no invented joint breakdown.
“Compare September with August.”	Explicit year; August baseline; absolute/relative count deltas, pp rate deltas and month-length caveat.
“Are our Google Ads campaigns sending more visitors?”	Evidenced paid/UTM filter and comparable trend; missing tags disclosed; no spend or ROAS claim.
“How many webinar registrations came from this page?”	Configured goal from get_goals, display_name used for activity; zero activity distinguished from missing configuration.
“Show traffic per week for the last six months.”	Exact bounds, weekly table and chart when supported; partial edge weeks and missing buckets labeled.
“Compare these two sites.”	Separate site queries and comparable scope/zones; no visitor sum presented as people.
“Maak hiervan een PDF-rapport.”	Actual Dutch PDF attachment with checked values/layout, or explicit unavailable-export fallback.
“What pricing strategy should our company use?”	No activation without a Plausible website-data question.

Also exercise zero baselines, null/empty results, unconfigured goals, sparse conversions, pagination, metadata warnings and unavailable app access. Verify missing history is not zero-filled, defining filters survive errors, relative dates use the discovered site zone, and external feedback requires authorization. Compare the same values across all three output modes and inspect each PDF page.

Additional acceptance cases for discovery and reporting contracts:

Case	Acceptance criterion
List sites or verify the default	Use get_sites with the appropriate selection; follow next_after for a full list; use timezone_status and site-local today.
Discovery denied or unsupported	Configuration stays unknown; no activity-based inference or request for elevated credentials.
Goal absent on the first page	Follow the complete cursor chain before declaring it unconfigured; use display_name for statistics.
Multiple breakdown/conversion pages	Preserve scope and limit across next_offset; inspect consistent totals and warnings; the last page is not independently complete.
Missing/invalid row totals or no progress	Report unknown completeness even for a short/empty page; stop without looping or claiming exhaustiveness.
Imported history requested	Pass include_imports: true; distinguish confirmed, excluded and unknown coverage, independently for each comparison period.
Structured metadata	Inspect preserved meta and query, including unknown fields, null metrics and empty results; no duplicate queries just to recover comparison metadata.
Prefix versus substring scope	Verify /traineeship* excludes /archive/traineeship; use an explicit contains filter for substring scope.