A Model Context Protocol (MCP) server exposing docs-search, summarize, and utility tools + resources over stdio, with a local demo client. Runs with zero keys.
Four tools with complete schemas and solid descriptions (avg 180 chars, within baseline 194). All parameters typed and described. Tool names follow verb_noun convention (search_docs, summarize, calculate, convert_units). Strengths: safe arithmetic implementation with clear constraints, good parameter validation (k clamped 1-20, max_sentences ≥1). Weaknesses: descriptions lack WHEN/WHY guidance for LLM selection; no output schema documentation; error handling present but not recovery-focused; no tool annotations (readOnlyHint); parameter descriptions could be more prescriptive (e.g., convert_units lacks unit enum or format spec).
Safely evaluate an arithmetic expression and return the result. Supports ``+ - * / // % **`` and parentheses over numbers only. It parses the expression to an AST and walks a strict allow-list, so it can never execute arbitrary code (no names, calls, attributes, or imports).
Convert a numeric value between compatible units. Supported dimensions: length (m, km, cm, mm, mi, yd, ft, in), mass (kg, g, mg, t, lb, oz), time (s, min, h, day, week), volume (l, ml, gal, qt, cup), data (b, kb, mb, gb, tb — binary/1024), and temperature (c, f, k). Common spellings/aliases (e.g. "meters", "pounds") are accepted.
Search the local developer knowledge base and return ranked snippets. Uses keyword/IDF scoring over the bundled markdown docs. Good for questions like "how does the stdio transport work?" or "what is semantic versioning?".
Summarize a block of text using the provider-agnostic LLM layer. Offline by default (a deterministic extractive mock); set ``LLM_PROVIDER`` to ``openai``, ``anthropic`` or ``ollama`` for a real, fluent summary.
Output schemas not documented. LLMs cannot infer what fields to expect from tool responses, forcing them to guess at downstream field names and types.
convert_units parameter descriptions lack format constraints. 'from_unit' and 'to_unit' accept free-form strings with no enum or pattern. LLMs will hallucinate invalid units (e.g., 'meter' vs 'm', 'kilogram' vs 'kg').
Tool descriptions lack WHEN/WHY guidance. 'Search the local developer knowledge base' does not tell LLMs when to prefer search_docs over summarize or calculate. No dependency hints or selection criteria.
No tool annotations (readOnlyHint, destructiveHint, idempotentHint). All four tools are read-only and idempotent, but this is not declared in the schema, forcing LLMs to infer safety properties.
Inferred effective spec: 2025-06-18+.
| Scored | Grade | Overall | Spec posture | Rubric |
|---|---|---|---|---|
| 2026-09-23 | D | 56 | 2025-06-18+ | v2 |
Error messages in calculate() and convert_units() are generic. 'Invalid status: got "pending", must be one of: open, in_progress, resolved, closed' pattern not followed. LLMs cannot self-correct without explicit constraint feedback.