amplitude

Use this skill for read-only product analytics through Amplitude's hosted MCP server.

1. Define the question before querying:

• Identify the metric or analysis type: active users, event segmentation, funnel, retention, chart, dashboard, experiment, cohort, taxonomy, session replay, feature flag, guide or survey, feedback, agent analytics, or user activity.
• Preserve explicit event names, property names, chart IDs, dashboard IDs, experiment IDs, cohort IDs, and user IDs exactly.
• Use the user's explicit time range. For relative ranges, state the concrete dates in the answer.
• If the request has no time range, use the shortest conventional range that answers it and state the assumption.

2. Discover the current Amplitude tool:

• Inspect the available Amplitude MCP tools for the required analysis.
• Use the exact live tool name and schema. Do not guess Amplitude MCP tool names or arguments.
• Prefer direct chart, dashboard, experiment, cohort, or user lookup when the user provides its identifier.
• If project or organization scope is ambiguous, inspect accessible scope through the read-only catalog and ask one focused question only when multiple plausible targets remain.

3. Query narrowly:

• Request only the events, properties, segments, metrics, and dates needed for the answer.
• Prefer one complete query over several overlapping queries.
• For comparisons, keep metric definitions and windows consistent.
• Stop once the requested result is supported; do not enumerate unrelated project metadata.

4. Report the result:

• Lead with the metric, trend, conversion, retention, or experiment finding.
• Include the date range, project or scope, filters, and metric definition needed to interpret it.
• Distinguish unique users, event totals, sessions, and percentages explicitly.
• Include Amplitude links returned by the MCP server when available.
• State when data is empty, delayed, inaccessible, or insufficient rather than inferring a value.

• DAU, WAU, MAU: use active-user or event-segmentation tools with the requested interval and clarify whether the result is unique users or event totals.
• Event segmentation: preserve the exact event and property filters; report the aggregation and interval.
• Funnels: preserve ordered steps, conversion window, exclusions, and segment filters; report overall conversion and material drop-off steps.
• Retention: report the starting event, returning event, cohort interval, and retention horizon.
• Saved charts and dashboards: fetch by ID when available and summarize the returned definition and results without recreating the analysis from memory.
• Experiments: report status, variants, exposure window, primary metric, and result confidence exactly as returned.
• Session replay: narrow by time, user, or event before retrieving replay details; summarize only the interactions needed to answer the question.
• Feature flags: report definitions, variants, and configuration without exposing deployment API keys or other credential fields.
• Guides, surveys, and feedback: preserve source and date filters, distinguish processed themes from raw comments, and avoid returning unrelated customer text.
• Agent analytics: distinguish sessions, spans, conversations, quality metrics, cost, and latency; return transcript content only when the user explicitly requests it.
• User activity: use an explicit user identifier when possible and avoid exposing unrelated profile properties or event history.

• Read-only only. The plugin exposes a fixed allowlist of Amplitude search, retrieval, and query tools.
• Do not create, update, archive, delete, launch, stop, or otherwise modify Amplitude resources.
• If the user asks for a change, explain that this plugin is read-only and offer to inspect the current state or describe the change they would need to make.
• Do not weaken or bypass the plugin's tool filter, including by calling Amplitude APIs through shell commands.
• Treat event and user properties as potentially sensitive. Return only fields needed to answer the request.
• Never expose API keys, credential fields, or unrelated conversation and feedback content returned by read tools.
• Do not fabricate event names, property values, metric definitions, experiment conclusions, or missing data.
• On authorization failure, let Junior present Amplitude's OAuth flow. On permission failure, report that the connected Amplitude account cannot access the requested scope.