AI Assistant

Around thirty registered tools, in three scopes. Tools outside the current scope are never advertised to the model — which measurably improves tool selection and costs no tokens.

• Always — the Excel side: read and write grids, workbook structure, chart generation, Python execution, sheet-structure analysis, smart fill, semantic formatting, dataset reconciliation, time-series forecasting, cell formatting; plus knowledge-base search, member where-used, artifact security, the Neural Graph and its filter, agent memory, and diagnostics.
• When connected to EPM — forms, data, export slices, metadata, sync, approvals, rule code, save impact (what happens when you hit Save on a form), audit records, security context, the group directory, and exact dimension statistics.
• Write tools — staging security assignments and provisioning groups. Gated: see the safety model below.

• Exact counts — asked how many members a dimension has, the assistant does not count from the knowledge graph, which samples members and would be wrong by construction. It reads the exact local outline and returns a real number.
• Complete where-used — indexed across forms, rules, dashboards, reports, action menus and navigation flows, so change impact stops being a judgement call.
• Model picker and PDF export — choose the model for a conversation from the chat header, and export any conversation as a white-page PDF.

The assistant auto-discovers tools from connected MCP (Model Context Protocol) servers — including CaptainMCP — so its reach extends to whatever systems you expose. Tool discovery is dynamic: add a tool to your MCP server and the assistant can use it immediately.

MCP integration requires a Node.js runtime on the workstation and is entirely optional.

Tools that change state (delete, import, update…) are classified as destructive automatically and require your explicit approval before they run. If no approval handler is available, destructive calls are refused — never silently executed.

• EPM write tools sit behind an explicit "allow EPM write actions" toggle and the approval dialog. With the toggle off — read-only mode — they are withheld from the model entirely, not merely blocked.
• Group provisioning is verified after it runs: the assistant reads the group back before reporting success.
• Every mutating call, user- or AI-driven, is recorded with target, outcome and timestamp in Captain's Log and exportable to Excel.

• Ribbon → AI Assistant — opens the chat panel.
• Right-click → Analyze with AI Assistant — sends the current selection as context. See Right-Click Tools.
• From other tools — analyzers like the Insights Analyzer and Performance Investigator can hand their findings to the assistant for root-cause follow-up.
• Beside the Neural Graph — while the graph window is open it responds to the chat's filter tool, so "show me only the business rules" re-forms the canvas.