UI, assistants and MCP
Use the supported interface for each forecasting action
Forecasting actions are exposed in project Predictions/Forecasters and through project-scoped MCP tools. The current public /api/v1 registry does not expose forecaster, forecast-run, watchlist or prediction-evaluation routes. The platform's session-authenticated /api routes serve its own UI; they are not documented here as API-key consumer endpoints.
Interface differences
| Interface | What it does |
|---|---|
| In-app forecaster creation assistant | Prepares, validates and commits a reviewable operator draft after approval |
| Manual creation | Name, prediction kind, evidence query, horizon and advanced prompt; derives technical defaults |
| Forecaster Edit | Schemas, execution settings, model, output paths, outcome policy and release selection |
| Run prediction page | Invokes an existing forecaster for a subject/window |
| External connected assistant | Can call the available MCP tools after reviewing actual resources and arguments |
| Prediction Outcome panel | Records outcome, observation time and notes; advanced observed values/evidence IDs use MCP |
Do not tell a reader to paste a tool-call envelope into a normal UI field. A prompt describes intent; MCP arguments are a structured tool request; an advanced schema/config editor expects only its object.
MCP action map
| Task | Tools | Required permission |
|---|---|---|
| Discover/inspect forecasters | forecaster_list, forecaster_get | projects:read |
| Create/update/archive | forecaster_create, forecaster_update, forecaster_archive | projects:write |
| Run a prediction | forecast_run | queries:execute |
| Inspect predictions | prediction_list, prediction_get | projects:read |
| Cancel | prediction_cancel | runs:execute |
| Schedule subjects | forecast_watchlist_create/update/remove | projects:write |
| Inspect schedule | forecast_watchlist_list | projects:read |
| Record outcome | prediction_evaluation_create | projects:write |
| Read history/report | prediction_evaluation_list, forecaster_evaluation_report | projects:read |
Check discover_capabilities on the actual connection. Tools are gated by forecasting availability and authenticated scope. Write tools require unique idempotency keys; cancellation uses id for the prediction UUID and also requires confirm true. Resource IDs must belong to the project. Tool availability does not create an evidence query or grant external source access.
Equipment assistant request
Inspect the published Equipment forecast evidence query and its subject contract. Prepare Equipment failure risk for an unplanned mechanical failure causing at least one hour downtime within 7d, with probability and separate confidence. Show the query release, schemas and single-call budget before creating. Run P-101 only after the configuration is reviewed, then show the stored evidence and prediction record.The in-app creation assistant prepares the forecaster portion. Use the Run page or an external connected assistant for invocation. Advanced output paths and evaluation policies should be inspected after the derived creation defaults.
MCP connection scope
The examples in this section use a project-scoped MCP connection and omit projectId from forecasting tool arguments. On the workspace MCP endpoint, supply projectId from project_list for each project action. Inspect the schema discovered from your actual endpoint; workspace tool schemas add that field. The workspace is bound by authentication.
API availability
Ontology evidence setup has public definition, binding and query APIs, as documented in Ontology interfaces. That does not make a forecasting API available. Use supported MCP forecasting tools for programmatic operations until public routes exist.