Returns a curated overview of what this MCP server can do, or a deep-dive on a specific area. Call this WITHOUT a topic when the user asks 'what can I do?' / 'how do I get started?' / 'what is Sheepit?' (or the Spanish equivalents — '¿qué es sheepit?' / '¿qué puedo hacer?' / '¿qué herramientas tiene sheepit?'). Call WITH a topic when the user asks about a specific area (campaigns, destinations, dashboards, insights, feedback, credentials). Pass language: 'es' when the user is writing in Spanish so the returned content matches their language.

Returns a step-by-step recipe naming the exact tool calls to chain for a specific goal. Use when the user has a clear intent ('send a marketing email', 'analyze the signup funnel', 'wire a webhook' — or in Spanish 'enviar un email de marketing', 'analizar el funnel de signups', etc.). Pass language: 'es' when the user is writing in Spanish.

Returns the events GoaTech understands out of the box (so they appear in pre-built dashboards / templates without manual rework) merged with the project's own registered custom events. Call this BEFORE writing any new track() / client.track() / useTrack() callsite — if a canonical event covers what you're about to emit, use the canonical name (e.g. signup_completed, not UserSignedUp or signup_done). Customers also benefit: their custom events show up under customer_events so you can match the convention they've already established. Filter to one category with category to avoid context bloat.

List user groups in the current project. Cursor-paginated, with optional free-text search across key and name. Use this BEFORE group_create to confirm the group doesn't already exist.

Create a new user group in the current project. Group keys are snake_case slugs unique per project. After creating, add members with group_add_member, then reference the group from a flag rule via {field: 'user_group', op: 'in', values: ['<key>']}. Example use cases: 'dogfooders' (early testers see new dashboards), 'beta_users' (cohort of opted-in feature testers), 'internal' (team members + advisors), 'banned' (denylist for kill-switches).

Add a user to a user group, by user UUID OR email. Provide exactly ONE of user_id / email — the server resolves email to user_id and 4xxs if no user has that email. Returns the membership row id. Idempotency: re-adding the same user returns 409 ALREADY_MEMBER. Adding to an archived group returns 409 GROUP_ARCHIVED.

Remove a user from a group by group id + user id. Returns 404 if the user wasn't a member. Idempotent in spirit — caller can treat 404 here as 'already gone' rather than an error.

List campaigns in the current project. Supports cursor pagination, status filter, free-text search across name/key/goal.

Fetch a single campaign by id with full audience / channels / creative / metric / budget / schedule / status.

Create a new campaign in draft status. The campaign isn't running until you call campaign_preview followed by campaign_launch. Channels: each entry needs a kind (email|meta|google|tiktok|linkedin|webhook). Audience: array of RuleCondition {field, op, values[]}. Success metric: {event_name, window_seconds?} — defaults to a 7-day conversion window.

Patch a campaign. ALLOWED only in draft or paused state. Trinary semantics for nullable fields: omit a field to PRESERVE its current value, send null to CLEAR, send a value to SET. Editing audience / channels / creative will INVALIDATE any outstanding preview_token (re-preview to get a fresh one).

Dry-run the campaign and return the launch plan + a single-use preview_token. The token is bound to the current state of audience/channels/creative/metric/budget/schedule and ALL of these must be present and valid for the token to be returned. Pass the token to campaign_launch within 5 minutes — editing the campaign in between invalidates it.

Move a campaign from draft|paused → scheduled|running. REQUIRES a fresh preview_token from campaign_preview. The token is single-use and snapshot-bound — if anything changed since preview, re-run campaign_preview to get a new token.

Pause a scheduled or running campaign. Pause is reversible via campaign_resume.

Resume a paused campaign (paused → running). Re-validates launchability — empty audience / channels / creative / metric will reject.

Move a scheduled / running / paused campaign to completed. Terminal — only archive follows.

Archive a completed (or never-launched draft) campaign. Removes it from the default list view but keeps history.

Fetch the latest results snapshot for a campaign. v1 only stores the latest aggregate (impressions, clicks, conversions, by_channel, by_variant); time-series snapshots land with the destinations framework.

Read-only enumeration of every destination adapter the server currently knows about (webhook, resend, etc.). Use this to discover which connector_id values are valid for destination_create. Returns id / version / category / title / description per adapter.

List destination configs installed in the current project. Filter by connector_id (e.g. only Resend installs) or status (active|paused|failed).

Fetch a single destination config by id with the full saved config + filters + last delivery state.

Install a destination adapter into the current project. connector_id must come from destination_catalog (e.g. "webhook", "resend"). config is per-adapter — the server validates it against the adapter's own schema and 400s on shape errors. Webhook config: { url: "https://...", signing_secret?: string, timeout_ms?: number }. Resend config: { from: "Name <addr@domain>", reply_to?: string, audience_limit?: number, batch_size?: number }. Returns the destination id — pass it as channel.destination_config_id on a Campaign to bind explicitly.

Update name / config / filters / status of an existing destination. Status transitions allowed: active ↔ paused (the "failed" state is system-set after consecutive delivery errors and cannot be set manually). Trinary semantics for nullable fields: omit = preserve, send a new value to overwrite.

Soft-delete a destination — it stops being eligible for campaign dispatch but its history (audit log of past deliveries) is retained.

Run the adapter's connection check against the saved config. For webhooks: GETs the URL to verify it's reachable. For Resend: lists domains with the API key to verify auth. Does NOT send a real campaign payload.

List every dashboard installed in the current project (excludes archived). Returns id / name / description / created_at per row. Use dashboard_get for the full widget list.

Fetch a single dashboard by id, including the full widget list (each with its query, viz, and position). Use this to understand what's on a dashboard before editing it.

Create an empty dashboard with name and optional description. Returns the new id. Use widget_create afterward to add widgets, or use dashboard_template_get + a script of widget_create calls to materialize a template manually. If you want a fully-populated starter, use dashboard_template_list to see what's available, then dashboard_template_get to get the widget specs.

Update name / description / layout of an existing dashboard. Trinary semantics for nullable fields: omit = preserve, send null to clear, send a value to overwrite. Templates (is_template = true) can't be edited via this tool — they're read-only.

Soft-delete a dashboard — it stops appearing in dashboard_list but its history is retained. Templates can't be deleted.

Read-only enumeration of every starter dashboard the platform ships with. Returns id / name / tagline / description / icon / widget_count per template. Use dashboard_template_get(id) to retrieve the full widget specs, then materialize via dashboard_create + widget_create calls.

Returns the full blueprint for a template (id, name, description, icon, full widgets array with each query + viz + position). Use this to fetch the recipe, then call dashboard_create + a widget_create per item to materialize it. Returns 404 if the template id is unknown.

Create a new widget on the given dashboard. type is a widget type (V1: 'timeseries' only). query is an InsightsQuery — a discriminated union on kind (V1: 'timeseries' only). viz controls presentation (chartType: 'line' | 'bar' | 'area' | 'stacked_bar' | 'single_metric'). position defaults to a sensible {x,y,w,h} if omitted. Templates can't have widgets added via this tool.

Update name / description / position / query / viz of a single widget. Trinary semantics for nullable fields. For bulk position changes (drag/drop save), prefer the dashboard layout endpoint at the API level — this tool is for one-widget edits.

Hard-delete a widget. The dashboard remains; only this single widget is removed.

Execute a one-shot InsightsQuery without saving it as a widget. Use this to answer questions like 'how many signups yesterday?' / 'errors-per-hour by app version this week?' / 'which utm_source converted best in the last 30 days?'. Input envelope: { environment_id?: uuid, query: { kind, event, interval, range, filters?, breakdownProperty?, aggregation? } }. query.kind is always 'timeseries' (the v1 surface). query.event is an event name from event_catalog_canonical. query.interval is one of 'minute'|'hour'|'day'|'week'. query.range is either {kind: 'relative', last: '1h'|'24h'|'7d'|'30d'|'90d'} or {kind: 'absolute', fromIso: iso, toIso: iso} (note: the absolute keys are fromIso/toIso, both full ISO-8601 datetimes). query.filters is an array of {field, op, value}; field names are dot-paths under event_properties / event_context (e.g. 'event_properties.course_slug'). query.breakdownProperty is a single property path that splits the series (caps at 20 distinct values). query.aggregation is {kind: 'count'} (default) or {kind: 'count_distinct', field: 'user_id'}. Returns gap-filled buckets; a missing time bucket is rendered as 0.

List releases newest-first with each release's CURRENT pre-computed health verdict (healthy/degraded/critical/unknown) and crash-free + error + latency rates. The verdict is computed server-side — read health_status, do not recompute it from the rates. Filter by platform, channel, or status (status='active' means deployed to production). Supports cursor pagination: pass cursor from pagination.cursor to fetch the next page. Use release_health for the full four-signal breakdown of one release, and release_regressions to find releases that got WORSE.

Return the full pre-computed health verdict and all four signals (crash-free, error rate, p95 latency, conversion) for ONE release, plus code context (commit SHA, PR number, PR title, branch, commit count) and the delta vs the prior release. health_status and health_score are authoritative — narrate them, never derive your own verdict from the individual rates. conversion_rate may be null when business metrics are not instrumented for this project; do not treat null as 0%.

List releases that REGRESSED (crash-free rate dropped past the server's significance gate), newest-first, optionally since a timestamp. kind='release' means the release degraded vs the prior release; kind='change' means it degraded right after a specific flag/config change (see change_entity_key). The drop is detected server-side — report crash_free_delta as given, do not recompute it. crash_free_delta of 0 on a kind='change' entry may mean the delta was not computable — check pre/post_crash_free_rate to confirm before treating 0 as a true no-change reading. Use release_health(release_id) to investigate a specific regression in depth.

File a bug report, feature request, or general note for the Sheepit team. The friction barrier between 'this is annoying' and 'report filed' is one tool call — use it. Call this proactively when the user expresses frustration ('this is confusing', 'I wish I could…', 'it should…'), when a tool returns a confusing error, or when you hit an obvious gap (a missing connector, a missing widget type, an unclear field name). Always confirm with the user before calling — quote their words back so the message is their voice, not yours. Returns the feedback id and createdAt timestamp. The MCP auto-stamps source/version metadata; you only supply type + message.