Orbis Prism MCP provides a comprehensive set of 20 tools for analyzing the Hytale Java API with consistent naming patterns (all start with prism_ verb) and reasonable descriptions. However, the server has significant gaps: output schemas are completely undocumented, parameter descriptions are minimal (typically 1 - 2 lines with no range constraints or format specifications), and error handling is absent from visible tool definitions. All tools are READ_ONLY and carry no risk classification beyond that. The codebase shows backend implementations with proper error handling (see index_queries.py returning (data, error_dict) tuples), but this error structure is not exposed in the MCP tool definitions, LLMs cannot see recovery guidance. Tools follow a clear naming convention (prism_<action>) which is strong, but lack the depth and structure expected of production-grade agentic tools.
Detect common software design patterns in a class
Get explanation of a Hytale API concept from the knowledge base
Find classes that implement or extend a target class
Find ECS systems that process a specific component
Find all usages of a class or method in source code
Get detailed information about a specific asset
Analyze who calls a specific method, grouped by package and class
Get detailed information about a specific class including methods and fields
No output schemas documented for any tool. LLMs cannot predict the structure of responses, forcing them to parse unstructured output. This violates pattern:tool and pattern:response-shaper. Without documented return types, agents cannot plan downstream tool calls or validate that required chaining IDs (e.g., package+class_name for follow-up queries) are present.
| Scored | Grade | Overall | Spec posture | Rubric |
|---|---|---|---|---|
| 2026-09-21 | D | 51 | 2026-07-28+ | v2 |
| 2026-03-09 | F | 26 | - | v1 |
Get list of indexed versions and active version
Get class hierarchy including parent classes and interfaces
Get statistics about the indexed Hytale API
Get detailed information about a specific method
Get a code snippet around a target string in a source file
Extract and inspect the bytes of a specific asset file
List classes matching a package prefix
List events and subscriptions in the Hytale API
List packages or subpackages
Read decompiled Java source file with optional line range
Run FTS5 search on indexed Hytale classes and methods
Search Hytale game assets by name or metadata
Parameter descriptions are minimal and lack actionable constraints. E.g., 'Search term or FQCN' for prism_search.query does not explain valid FQCN format, length limits, or what happens on partial matches. 'Maximum results to return (1-500)' for limit is adequate but inconsistent, other tools omit the range entirely. Per pattern:constrained-input, numeric ranges should be explicit in the description.
Error handling is invisible in tool definitions. The backend code (index_queries.py) implements proper error classification (no_db, not_found) with recovery hints ('Did you mean...?' suggestions), but MCP tool definitions do not surface this. LLMs receive no guidance on retryability, root cause, or recovery steps. Per pattern:recovery-guide, error responses must tell the LLM what to do next.
Tool descriptions are between 34 - 72 characters (below the 50 - 200 char baseline for LLM optimization). Example: 'Get list of indexed versions and active version' (47 chars) does not explain WHY the LLM should call this tool, WHEN to use it, or what problem it solves. Per pattern:tool-description, descriptions should state context, purpose, and preconditions.
Version parameter is required on most tools but has no description explaining valid values, format, or default behavior. The backend normalizes versions ('release', 'prerelease') per domain/constants.py, but this contract is not exposed to the LLM. Per pattern:tool-description, parameter descriptions must specify allowed values (as enum or as text).
Pagination support is inconsistent. prism_list_classes exposes limit and offset; prism_list_packages, prism_list_events do not document pagination at all. The backend limits results (limit capped 1 - 500 in code), but tools accepting 'limit' as unlimited parameters invite LLMs to request thousands of results, risking context window explosion. Per pattern:paginated-result, all list tools should enforce and document max page size.