A Model Context Protocol server for Prem AI - enables AI assistants to interact with Prem's ecosystem for chat completions and RAG capabilities.
Three tools with visible Zod schemas and descriptions, but significant gaps in parameter documentation and output schema clarity. Tool names lack action verbs (chat, prem_upload_document, prem_chat_with_template). Parameter descriptions are present but generic, e.g., 'Optional temperature for response generation' lacks range/constraint guidance. No documented output schemas; responses are JSON-stringified API payloads rather than structured, agent-friendly formats. Error handling returns raw error messages without recovery guidance. Naming inconsistency: prem_upload_document and prem_chat_with_template use prefixes while chat does not.
Chat with Prem AI - supports chat completions with optional RAG capabilities.
Chat using a predefined Prem AI prompt template
Upload a document to a Prem AI repository
Tool names lack action verbs. 'chat' should be 'send_chat' or 'create_chat_completion'; 'prem_upload_document' is verbose and inconsistent with 'chat'. LLMs rely on verb-noun naming to infer intent.
Output schemas not documented. Tools return JSON.stringify(response) as plain text. Agents cannot parse structured fields or chain results. No documented return types for temperature, max_tokens, repository_ids parameters.
Parameter constraints missing. temperature, max_tokens, similarity_threshold, limit lack min/max bounds or valid ranges. LLMs will pass arbitrary values (e.g., temperature=999, max_tokens=-1).
Inferred effective spec: <=2025-11-25.
| Scored | Grade | Overall | Spec posture | Rubric |
|---|---|---|---|---|
| 2026-09-23 | F | 41 | <=2025-11-25 | v2 |
Error handling provides no recovery guidance. Errors return raw messages like 'Chat error: <message>' or 'Document upload error: <message>' without actionable next steps or error classification (retryable vs fatal).
Parameter descriptions are generic and lack format/constraint details. E.g., 'Optional temperature for response generation' does not specify range (0.0 - 2.0?). 'Optional array of repository IDs for RAG' does not explain what repository IDs are or how to obtain them.