MCP server that exposes a content pipeline for AI agents to research topics, generate summaries, and create YouTube scripts
Three tools with basic verb-noun naming (search_topic, summarize_research, generate_youtube_script). All have descriptions (10-60 chars, below the 194-char baseline). Input schemas present but minimal: only 1-2 parameters per tool, all strings with basic descriptions. No output schemas documented. No error handling guidance, no parameter constraints (enums, ranges), no idempotency hints. Tools lack the depth needed for reliable agent composition, e.g., generate_youtube_script chains three internal calls but doesn't expose intermediate results or recovery paths. No tool annotations (readOnlyHint, destructiveHint). Descriptions are too terse to guide LLM selection; 'Research a topic on the web and return a written summary' doesn't explain when to use summarize_research vs generate_youtube_script.
Research a topic and generate a short YouTube script from it.
Search the web for current information about a topic via Tavily.
Research a topic on the web and return a written summary.
Descriptions are too short (10-60 chars vs 194-char baseline). 'Search the web for current information about a topic via Tavily' lacks context on when to use search_topic vs summarize_research vs generate_youtube_script. LLMs cannot disambiguate.
No output schemas documented. Callers don't know what fields to expect. search_topic returns dict, summarize_research and generate_youtube_script return str, but no structure for dict keys, no guidance on downstream tool chaining.
No error handling guidance. If Tavily API fails, research_topic() will raise an exception with no recovery hint. LLM receives a stack trace instead of 'Tavily search failed. Try a simpler query or check API key.'
| Scored | Grade | Overall | Spec posture | Rubric |
|---|---|---|---|---|
| 2026-09-23 | F | 44 | 2026-07-28+ | v2 |
generate_youtube_script chains three internal calls (research_topic → generate_summary → generate_script) but exposes no intermediate results. If summary generation fails, the LLM cannot retry just that step or inspect partial results. No idempotency hints.
tone parameter in generate_youtube_script has a default ('Conversational') but no enum constraint. LLM could pass 'Sarcastic' or 'Angry', backend may not support these. Should declare enum: ['Conversational', 'Energetic', 'Informative', 'Humorous'] (matching Streamlit UI).