Server has 7 tools with explicit FastMCP registration and visible schemas. Naming follows verb_noun convention (cnki_login_status, cnki_search, cnki_download_pdf). Descriptions are present but vary in quality: some are detailed (cnki_batch at ~400 chars), others minimal (cnki_save_cookies at ~80 chars). All tools have input schemas with type definitions. However, critical gaps: (1) No output schemas documented, LLMs cannot predict return structure; (2) Parameter descriptions lack constraints (e.g., year_start/year_end have no min/max bounds, db_code has no enum); (3) No error handling guidance, tools return dicts but no recovery hints; (4) STDIO transport caps score at 50 hard limit. cnki_batch is complex (7 params, irreversible) but lacks confirmation/dry-run pattern.

• Add explicit output schemas to all tools. Define return type as JSON Schema with properties (e.g., {type: object, properties: {success: {type: boolean}, papers: {type: array, items: {type: object, properties: {title, authors, year, journal, url}}}}}). This enables LLM to plan downstream calls.
• Convert db_code to enum: {type: string, enum: [CJFD, CDFD, CMFD], description: 'Database code: CJFD=journals, CDFD=PhD theses, CMFD=Master theses'}. Add min/max to year_start (1900), year_end (2100), max_results (1 - 100).
• Add error recovery hints to tool descriptions. E.g., cnki_download_pdf: 'If CAPTCHA triggered, browser stays open, complete verification manually. If network timeout, retry with longer captcha_wait. If PDF unavailable, check journal access permissions.'
• Implement dry-run for cnki_batch: add optional parameter 'dry_run: boolean = true' (default safe). When true, return preview of papers to download + Zotero import count without executing. Agent can review then call with dry_run=false to commit.
• Migrate from STDIO to HTTP+SSE or Streamable HTTP transport. Use FastMCP's built-in HTTP support (mcp.run_server(transport='sse') or 'http'). This enables remote deployment and hosted client compatibility.