MCP server for searching products in Bonsai's Elasticsearch cluster
The Bonsai MCP server has 1 tool (search) with significant quality gaps. The tool has a basic description (47 chars) that meets minimum length but lacks depth. The input schema is minimal, only a 'query' parameter with a basic string type and description. There is no documented output schema, no error handling guidance, no parameter constraints (e.g., min/max length), and no recovery hints. The tool name 'search' is generic and does not start with a clear action verb pattern (e.g., 'search_products'). No input validation, pagination support, or result limiting is evident. The response is returned as a raw JSON string from format_response(), not as structured data. This server exhibits patterns typical of a minimal proof-of-concept rather than production-grade tooling.
Search for products in Bonsai's cluster
No documented output schema. The search tool returns a formatted JSON string, but callers do not know the structure of the returned object. LLMs cannot plan downstream operations or extract specific fields reliably.
Missing input parameter constraints. The 'query' parameter accepts any string with no documented limits on length, allowed characters, or format. LLMs may pass excessively long or malformed queries that fail silently or cause API errors.
No error handling or recovery guidance. If the Bonsai API fails, returns no results, or encounters a network error, the tool returns generic strings like 'No response received' or 'No search results found'. LLMs receive no actionable guidance on what to do next (retry, refine query, check configuration).
| Scored | Grade | Overall | Spec posture | Rubric |
|---|---|---|---|---|
| 2026-09-22 | F | 32 | 2026-07-28+ | v2 |
| 2026-03-09 | F | 31 | - | v1 |
No result limiting or pagination. The tool returns the entire raw JSON response from Bonsai, which could contain hundreds or thousands of hits. No pagination, limit, or offset parameters are offered, risking context window exhaustion and token waste.
Generic tool name 'search' lacks specificity. The rubric baseline shows most production tools start with action verbs like 'get', 'list', 'create', 'search'. A more descriptive name like 'search_products' or 'search_bonsai_catalog' would help LLMs disambiguate when multiple search tools are available.
Unhandled API failures. The make_search_request() function does not catch exceptions (e.g., network errors, JSON decode failures, HTTP 500s). Callers receive no indication of what failed or how to recover.
Credentials exposed via environment variables but no validation. The BONSAI_URL and INDEX_NAME are loaded from environment but not validated at startup. Missing or malformed config will only fail at tool invocation, wasting LLM steps.
No tool annotations. The search tool does not declare readOnlyHint, destructiveHint, or idempotentHint. The tool is read-only, so at minimum it should be marked as such for agent safety.