Skip to content

Resources & prompts

In short

  • Four parleml:// resources: a browsable index, the playbook topics, AutoDoc model cards and workspace catalogs. Each has a tool equivalent.
  • Seven prompts scaffold common jobs, such as audit_model and deliver_model. Hosts show them like slash commands.
  • 17 ui://parleml/… app templates render tool results as panels (Apps).

Beyond the tools, the server exposes four parleml:// resources and seven prompts. Hosts surface these differently — a resource picker, a slash command, an attachment menu — so they are worth knowing even though every one has a tool equivalent.

Resources

parleml://index is the only listable resource: hosts that show a resource browser show this one, and it links onward to everything else. The other three are templates — the host (or the model) fills the {…} segment.

URI MIME Same data as
parleml://index text/markdown — (a browsable index of the rest)
parleml://guide/{topic} text/markdown get_playbook(topic)
parleml://autodoc/{experiment_id} text/markdown get_autodoc(experiment_id)
parleml://catalog/{workspace_id} application/json describe_catalog(workspace_id)

Valid {topic} values are the same set get_playbook accepts: overview, workflow, api, errors, charts, explainability, schema-{supervised,unsupervised,forecast,recommendation,association}, and the builder workflow's deliver, iterate and byo, plus type-supervised, type-forecast, type-unsupervised, type-recommendation, type-association and type-text. An unknown topic is a ResourceError, not a success payload carrying an error sentence — so a client can tell "here is the guide" from "no such topic".

Resources go through the same auth, error mapping and per-tool telemetry as the tools; a failed read surfaces as a ResourceError with the same actionable hints (401 → mint a token, 404 → publish results first).

Prompts

Seven prompts scaffold the common jobs and expand to a short plan the model can follow or adapt. All seven are registered for prompt discovery; the first four guide analyst jobs (with compute required for optional probes), and the other three guide work whose tools need a builder token. Each names its STOP points (where the agent hands back to the user) and ends at a person's decision. Every prompt opens with the same rule: ParleML computes, the agent reasons — it never re-implements a number a tool provides.

Each prompt takes names as the user would say them; the scaffold starts by resolving them to ids.

Prompt Arguments Scaffolds
explore_workspace(workspace) workspace name catalog → profile → first questions
audit_model(experiment) experiment name evaluation → trust / fairness / drift → rules
investigate_anomalies(experiment) experiment name score distribution → top-k → per-row drivers
triage_dataset(dataset) dataset name EDA → leakage → optional probe_dataset (signal, baseline floor)
deliver_model(workspace, question) workspace, the business question brief → frame (STOP) → recipe → prepare → plan (STOP) → launch → iterate → decide → register
prepare_training_table(workspace, question) workspace, the business question catalog → target and prediction moment (STOP if ambiguous) → recipe SQL → preview → materialise → roles → leakage triage → signal
improve_experiment(experiment) the run to improve on one lever per experiment (STOP before each launch) → compare vs parent and the model in production → register a winner for review

MCP-App panels

When the frontend embed build has run, the server also serves seventeen ui://parleml/<id> App templates — evaluation, eda, rules, catalog, chart, insight, resource, run, signal, compare, decision, what-if, plan, recipe, prepare, scoring and model-card — hosting 51 panels. Each linked tool carries the standard _meta.ui.resourceUri, so a host that supports MCP-Apps renders the app for that tool's results automatically. An app with several panels picks one from a field of the result (chart_kind, kind, resource_kind, view, probe, section, compare_kind, simulation, action — see Apps).

Panels are sandboxed HTML with a strict CSP (default-src 'none', connect-src 'none', form-action 'none'), and each ui:// resource declares the MCP Apps CSP with no network origin at all — they render the payload the tool already returned and make no network calls of their own. A build without the embed step simply serves no panels; every tool still works.

Next