MCP (Model Context Protocol) server library for Elixir with per-session GenServer architecture
NexusMCP demonstrates solid foundational structure with 12 tools, all with descriptions and input schemas. Tool naming follows verb_noun convention (greet, get_status, add, delete_item, get_weather, etc.). However, descriptions are uniformly terse (10-30 chars), falling well below the 50-200 char LLM-optimized baseline. Parameter descriptions exist but are minimal ('Person's name', 'City name'). No output schemas are declared for any tool, violating the documented output schema pattern. Error handling is basic, fail_tool and raise_tool exist for testing but production error guidance is absent. The deftool macro supports annotations and output_schema, but these are not utilized in the 12 tools shown. Tool composition is reasonable (single responsibility), but parameter constraints (enums, ranges, formats) are not visible in schemas.
Add two numbers
Get audit log entries
Check process dictionary state
Delete an item
Echo a message
Tool that fails intentionally
Get current status
Get current weather
All tool descriptions are under 30 characters, far below the 50-200 char LLM-optimized baseline. Descriptions like 'Get current status' and 'List items' lack context for when/why to use the tool and what it returns.
No output schemas declared for any of the 12 tools. The deftool macro supports output_schema parameter, but it is not used. LLMs cannot plan downstream calls or extract structured data without knowing return field types.
| Scored | Grade | Overall | Spec posture | Rubric |
|---|---|---|---|---|
| 2026-09-23 | D | 59 | 2025-06-18+ | v2 |
Greet someone by name
List items
Tool that succeeds
Tool that raises an exception
Parameter descriptions are minimal (e.g., 'Person's name', 'City name', 'Item ID'). No format constraints, ranges, or enums visible. LLMs cannot infer whether 'city' accepts 'New York' vs 'NY' vs coordinates.
list_items and audit_log lack pagination parameters (limit, offset, cursor). Returning unbounded lists risks context window exhaustion and LLM reasoning degradation.
Error handling is minimal. fail_tool and raise_tool exist for testing, but production tools lack recovery guidance. No error categorization (retryable vs user-fixable vs fatal) visible.