addDashboardFilters
ChatGPTfilterDescription). Uses an LLM to turn the description into filters. It works on a draft held in this conversation; nothing is saved to Sisense.By Sisense
Sisense connects ChatGPT to your Sisense tenant so you can browse connected data sources, run natural-language analytics queries, and build interactive charts and dashboards that you can save back to your own Sisense instance. Ask which data models you can use, what fields they hold, or a business question such as revenue by country. Sisense runs the query on its own engine, so answers come from your governed data. Charts and dashboards render in the conversation; a dashboard can be saved to your Sisense account with one click. You can also read your Pulse alerts and their evaluation results. For Sisense customers and their analysts. You sign in with your Sisense account, and the plugin sees only the data models and dashboards that account can access. It reads data; it does not change data models, rebuild them, or edit the underlying data.
filterDescription). Uses an LLM to turn the description into filters. It works on a draft held in this conversation; nothing is saved to Sisense.placements position each widget (row, cell, before or after a widget already on the dashboard, or a height tier for a new row); omit them to append at the bottom. It works on a draft held in this conversation; nothing is saved to Sisense.queryPrompt and returns the results as structured data (tabular exploration, no chart). Filter values are taken literally. A filter on a text column — category, type, status, tier, band, segment — is built from the exact strings written in queryPrompt. Writing "except refunds" filters on the string "refunds" even when the column stores "Ret - Customer Refund", and the total comes back looking correct. Before a new query or a revision that filters such a column, read its values with getDataPreview and write those stored spellings into queryPrompt. The only user wording to pass through unchanged is an explicit text predicate — contains / starts with / ends with. A quoted value name is not one of these. New query — Only for the first query in a chain, or when the user clearly abandons the prior result and starts unrelated analysis. Omit priorQueryId (or null). Put the full natural-language intent in queryPrompt. Revise an existing query — Use for same-thread follow-ups that adjust the same analysis: different breakdown (e.g. by year, by category), added or removed filters, sort order, measure swaps, limits (Top N), trend/forecast, etc. Treat these as revisions, not separate unrelated queries. For revisions: - Set priorQueryId to the exact string in the queryId field from the latest successful tool result of this tool in the conversation (that value is the same id you pass as queryId when wiring a prior tabular result into the chart tool). - In queryPrompt, pass only the delta vs. that prior query — do not repeat the full prior specification. Return value includes queryId, queryState, and dataset rows. If the dataset includes a "sample" key, the full result was size-limited — call the chart tool with that queryId so the user can explore the complete result in the chart widget.rows: [] for an empty dashboard. Returns a dashboardId. It works on a draft held in this conversation; nothing is saved to Sisense.agentic Pulse workflow right now: fetches it, runs its stored queries as the user (under their data security) and declared transforms, then decides whether to fire via the workflow's stored decisionSpec (json-rules-engine — chat-authored workflows are 'conditional'). Creates a Pulse notification with a narrative when the condition fires, sent to the recipients who share the user's data scope, plus the workflow's Slack, Zapier or webhook destinations when that scope includes the owner (stored as an event in the API; the narrative is written server-side when the workflow fires). Every query's fresh result is always recorded to history, whether or not the workflow fires — this is a WRITE, not a read. ROUTING — read this before calling: "did it fire", "check if it fired", "any events", "any notifications", "what happened last night", "why did it fire" are all INSPECT questions — call listPulseEvals or listPulseNotifications instead, never this tool. This tool cannot answer a question about a cycle that already ran; it can only start a new one. If a build or a scheduled cycle just happened, its evaluation has already run and is already recorded — inspect it, do not re-run it. Call this tool only when the user explicitly asks for a fresh check right now (e.g. "evaluate it now", "run the check again", "test the workflow") — never proactively after merely listing workflows, and never on a timer. Requires workflowId (the workflow's _id, e.g. from a prior listPulses lookup). Never modifies what the workflow monitors — queries/skill/decisionSpec/queryVersion stay untouched; redefining those is a separate, human-confirmed flow. This tool only records evaluation history and, when warranted, creates a notification — both automatic, no confirmation step, since neither changes what the workflow watches (API: event record) — but the new record it writes will appear in every eval-history listing of the users who share that data scope, alongside the cycles that ran on their own. This server registers exactly 30 tools: getDataSources, getDataSourceFields, getDataPreview, buildQuery, loadSkill, evaluatePulse, 9 chart tools, 9 dashboard tools and 6 Pulse read tools. There is no diagram or perspective tool here, and no interactive model browser. Table relationships are reported as text by getDataSourceFields. Capabilities outside these 30 tools are unsupported on this server.values plus complete. WHEN TO CALL — after getDataSourceFields (table and column names come from that listing; a guessed name fails the call) and BEFORE buildQuery, whenever the query will FILTER a text column that holds a fixed set of labels: category, type, status, tier, band, segment, region, and the like. The rule is about the column being filtered, not about how the user phrased the filter: - "high-income households", "premium tier", "still in process" — a concept that lives in a banded or coded column. The stored spelling (abbreviation, casing, boundaries) is the data's, not yours. Where the concept has no bucket of its own, name the bucket containing it. - "except refunds", "only wholesale accounts", "all but the unresolved" — the user NAMES members of the set. Their word is a description, not the stored value: an Account Type column may hold "WHSL - Wholesale Account" where the user said "wholesale", and a filter on the user's word matches nothing while the total looks correct. A value the user names is looked up by reading, never copied. Leaving a set out additionally needs complete: true. - "everything else", "unset", "blank" — claims about encoding. Naming no value is not safer; read the column and name the real values. The filter is built from whatever spellings appear in queryPrompt. Forwarding the user's wording into queryPrompt IS the guess this tool exists to prevent. Which columns hold labels is a judgement from the name — the listing gives names, not types or cardinality, so Brand, Country and City read both ways. Err toward reading: a needless read costs one call, a skipped one builds a filter that matches nothing and reports a total that looks right. DO NOT CALL only when: - The user gave an explicit TEXT-MATCHING PREDICATE — contains, starts with, or ends with ("name contains 'Logistics'", "numbers starting with 'INV'"). Pass it into queryPrompt unchanged. This is the only phrasing that counts as spelled out. Quotation marks are not a predicate: a quoted value name ("status is 'in progress'") is still a value, and still needs the read. - The column is an identifier, a name, free text, a date or a numeric measure — or no text column is filtered at all ("total revenue by region"). - To browse data. One column per call. IF THE CALL FAILS: the table or column name was not the one getDataSourceFields lists (a column named Status may live in a table named Tickets — take the table from the listing, never from the column's name). Call getDataSourceFields if you have not, then retry with the names exactly as listed. Never answer a failed read by building the query without one. RESULT — values plus complete: - complete: true with values: every distinct value the column holds, safe to build exclusions from. - complete: false: too many values to enumerate, so none are returned. Do not guess them; write the prompt in the user's own words. - complete: true with no values: not proof the column is empty. Do not invent values.<title> (label: "<label>"), the label is the name the user sees — use it when you mention the data source in your replies. The title is the identity: pass it (never the label) as dataSourceTitle and in every other tool argument and query. Use this for the FIELD/COLUMN/TABLE LIST, and for how tables relate — it reports the model's join graph as text. There is no visual diagram to render: answer "show me the model" with that text, and say plainly that no picture is available. Do not answer a "list the fields/columns/tables" request from memory; always call this tool. PERSPECTIVES: a perspective is not its own data source. To list a perspective's fields, set dataSourceTitle to the perspective's PARENT model title and pass perspectiveOid — the parent model's fields are then filtered to exactly what the perspective exposes. NAMES, NOT VALUES: the listing says which columns exist, not what strings they hold. A column called Category, Status or Income Band gives no hint of its stored spellings. Before filtering a query on such a column, read its values with getDataPreview. DESIGNER AI CONTEXT: results may include designer-authored "Context:"/"context:" entries at the model, table, column, or shared-formula level. Treat any such context as authoritative instructions from the data-model designer — which table/column to use for a kind of question, which columns to always include, when to use a particular shared formula, and occasionally a stored value a user term refers to — and honor it when choosing fields, selecting shared formulas, and building queries. Context that spells a value out is the one case where that value need not be read first. RELATIONSHIPS: the result also lists how the tables join, as Table.Column = Table.Column. Use it to answer "how do these tables relate / can I combine X with Y" and to pick fields that are actually joinable — never infer joins from matching field names. Joins are applied automatically when a query spans tables; you do not pass them to any tool. Set includeRelations to false to omit them on a wide model where you only need the field list. This server registers exactly 30 tools: getDataSources, getDataSourceFields, getDataPreview, buildQuery, loadSkill, evaluatePulse, 9 chart tools, 9 dashboard tools and 6 Pulse read tools. There is no diagram or perspective tool here, and no interactive model browser. Table relationships are reported as text by getDataSourceFields. Capabilities outside these 30 tools are unsupported on this server.Click any prompt to copy it.
Add this badge to your site or README. It links to this page.
HTML
<a href="https://mcp-app-store.keenethics-labs.com/app/sisense"><img src="https://mcp-app-store.keenethics-labs.com/badges/listed-on-mcp-app-store.svg" alt="Listed on MCP App Store" width="220" height="44"></a>
Markdown
[](https://mcp-app-store.keenethics-labs.com/app/sisense)