Canvas LMS Model Context Protocol with lazy authentication for Smithery integration
This server exhibits significant structural and documentation gaps. While it attempts to implement basic MCP tool semantics, it lacks proper schema definitions, adequate parameter descriptions, and error handling guidance. The two tools (get-courses, get-assignments) are minimally documented and lack the structured output schemas required for LLM comprehension. Tool naming follows a reasonable verb_noun convention, but parameter documentation is sparse, and no output schemas are visible in the source. The 'lazy authentication' pattern, while interesting, is not a substitute for proper tool definition and is not documented in tool descriptions.
Retrieves assignments for a specific course
Retrieves courses from Canvas LMS
No output schemas documented. Tools define input parameters but provide no specification of response structure. LLMs cannot reason about what fields to expect or how to chain tools.
Tool descriptions are trivially short (15-18 chars). 'Retrieves courses from Canvas LMS' and 'Retrieves assignments for a specific course' lack context for when to select these tools over alternatives, dependencies, and prerequisites. Baseline is 194 chars; these are <10% of baseline.
Parameter descriptions are missing context. 'Maximum number of courses to return' (limit) is adequate but lacks range constraints (min/max). 'ID of the course' (courseId) does not explain format or whether it accepts slugs/names in addition to system IDs.
| Scored | Grade | Overall | Spec posture | Rubric |
|---|---|---|---|---|
| 2026-09-22 | F | 35 | 2026-07-28+ | v2 |
| 2026-03-09 | F | 37 | - | v1 |
No parameter validation guidance in descriptions. The 'limit' parameter on get-courses has no documented min/max bounds, inviting LLMs to pass absurd values (0, -1, 100000). No enum constraints on courseId format.
Credentials (apiKey, domain) passed in request body as 'authentication' field. While lazy-loaded on execution, this pattern is non-standard and not documented in tool descriptions. The tool descriptions do not mention authentication requirements, forcing LLMs to infer them from execution errors.
No error handling guidance. Controller returns generic error responses ('An unknown error occurred', 'Tool with ID X not found'). No categorization of retryable vs user-fixable vs fatal errors. No actionable recovery suggestions.
No pagination support visible. get-courses accepts a 'limit' but no offset/page parameter, cursor, or total count in response. Large result sets will blow context windows without pagination.
Tool composition breaks LLM chaining. get-assignments requires 'courseId' but get-courses output schema is not documented. LLMs cannot determine if courseId matches the ID returned by get-courses without guessing.
No input validation in tool execution. executeTool() validates required params but does not validate types, ranges, or formats. A 'limit' of -1 or 'xyz' will silently pass to Canvas API.