MCP server for the ROLLIN Accessibility API — wheelchair accessibility data for 98,000+ US locations across 16 states
This server presents a significant paradox: the tool definitions themselves are reasonably well-crafted with descriptions and schemas, but the implementation contains aggressive trial-mode mechanics (5-request limit, sales pitches, session tracking) embedded directly in the tool handlers that will degrade the experience and violate agent autonomy principles. The four tools are READ_ONLY/WRITE operations on accessibility data, but the core tool definitions lack critical details: output schemas are not documented, parameter constraints are missing (e.g., 'limit' is unconstrained numeric), and error handling does not guide recovery. The server also embeds non-optional sales/upsell logic into every tool response during trial mode, which violates the principle that tools should serve user intent without behavioral manipulation. Naming is verb-forward and clear (search_locations, location_details, list_regions, submit_feedback), and descriptions exist for tools and most parameters. However, parameter descriptions lack actionable constraints (format, range, allowed values), and output schemas are entirely undocumented. The implementation also tracks session state (searches, locations_found, details_viewed, feedback_given) to personalize sales pitches, which couples tool logic to marketing strategy and will confuse agents attempting to reason about tool behavior.
List all covered regions/states where ROLLIN has accessibility data. Shows coverage area and update status for each region.
Get detailed accessibility information for a specific venue by ID: full feature checklist, community reviews, accessibility ratings, contact details, and recent updates.
Search for wheelchair-accessible restaurants, cafes, and bars near a location. Returns scored results with accessibility features, contact info, and ratings.
Submit accessibility feedback or corrections for a venue. Updates are reviewed and merged into the community-curated database.
Output schemas completely undocumented for all four tools. Agents cannot predict response structure, forcing them to guess field names and types. This violates the fundamental contract that agents need to know what data they will receive.
Trial mode mechanics (5-request limit, sales pitches, session tracking) embedded in tool handlers violate agent autonomy. Tools should not condition responses on licensing state or inject marketing messages into every call. This conflates tool logic with business logic and will confuse agents attempting to reason about consistent behavior.
Parameter 'limit' in search_locations lacks min/max constraints. Rubric requires: 'Specify minimum and maximum for numeric parameters (e.g. page_size 1 - 100)'. Current definition allows unbounded values that could stress downstream API.
| Scored | Grade | Overall | Spec posture | Rubric |
|---|---|---|---|---|
| 2026-09-22 | D | 55 | 2026-07-28+ | v2 |
| 2026-03-09 | F | 0 | - | v1 |
No pagination mechanism documented for search_locations. Description claims results are 'scored' and 'returned' (plural) but no limit, offset, page, or cursor parameters exist beyond the single 'limit' field. For queries spanning 98,000+ locations, result sets could be large; without pagination guidance, agents cannot compose multi-page searches.
Error handling provides no recovery guidance. Trial-expired responses and rate-limit nudges use structured JSON (good) but do not indicate whether errors are retryable, user-fixable, or fatal. Agents cannot determine next steps from error responses.
Session-state tracking (searches, locations_found, details_viewed, feedback_given) couples tool execution to user behavior analytics, violating separation of concerns. Tools should not track behavioral data for personalization; this belongs in application-level logging, not tool handlers.
Parameter descriptions for 'category' in search_locations mention 'optional category filter: restaurant, cafe, bar', these should be declared as an enum constraint, not free-form text. Rubric: 'When a parameter accepts one of a known set of values, declare it as an enum.'
submit_feedback description does not clarify whether feedback is applied immediately, queued for review, or rejected with validation errors. Agents need to know the outcome (success, pending, error) and any follow-up actions.