MCP-server for sanntidsdata fra Ruter/Entur — er bussen i rute?
Two tools with clear, domain-specific descriptions (Norwegian transit context). Both have complete input schemas with type definitions and parameter descriptions. However, naming lacks action verbs (finn_stoppested, avganger are descriptive but not verb-first), and output schemas are not formally documented. Error handling returns text-wrapped JSON but lacks recovery guidance or error classification. Parameter constraints are present (regex, min/max) but some descriptions embed examples that could be reused literally by LLMs.
Hent avganger fra et stoppested med sanntidsinfo: planlagt vs. forventet tid, forsinkelse i minutter og eventuelle avviksmeldinger. Svarer på «er bussen i rute?». Bruk NSR-id fra finn_stoppested (f.eks. «NSR:StopPlace:58366»). Kan filtreres på linjenummer og/eller destinasjon.
Søk etter holdeplass/stasjon i Norge etter navn (f.eks. «Jernbanetorget» eller «Nydalen»). Returnerer NSR-id-er som brukes i avganger-verktøyet. Bruk alltid dette først hvis du ikke kjenner id-en.
Tool names lack action verbs. 'finn_stoppested' and 'avganger' are descriptive but do not start with verb_noun pattern (e.g., search_stop, get_departures). LLMs rely on verb-first naming to infer intent before reading descriptions.
Output schemas not formally documented. Tools return JSON via somTekst() wrapper, but the structure of Stoppested[], AvgangerResultat is not exposed in tool registration. LLMs cannot plan downstream calls or extract fields reliably.
Error handling lacks recovery guidance. somFeil() returns raw error messages without actionable next steps. E.g., 'Entur svarte 503' tells the agent nothing about retry strategy or fallback options.
| Scored | Grade | Overall | Spec posture | Rubric |
|---|---|---|---|---|
| 2026-09-23 | C | 60 | 2026-07-28+ | v2 |
Parameter descriptions embed example values ('Majorstuen', 'Fornebu', 'NSR:StopPlace:58366'). LLMs frequently reuse examples literally in real calls, causing failures when the example does not match the user's intent.
No pagination support. avganger accepts 'antall' (max 20) but does not return a total count, next_cursor, or offset. Large result sets could exhaust context; no way to iterate.