priorwork-mcp

# priorwork-mcp MCP server for [priorwork.fyi](https://priorwork.fyi) — semantic search over 114,000+ papers from major ML, NLP, and Computer Vision conferences (NeurIPS, CVPR, ICML, ICLR, ACL, EMNLP, ICCV, AISTATS, and more), directly from Claude Code, Cursor, or any MCP client. ## Tools | Tool | Description | |---|---| | `search_papers` | Semantic search from a text query (a short descriptive sentence or two works slightly better than a few keywords) — returns ranked papers with authors, venue, year, similarity score, download URL, and full abstract. Pass `summarize=true` to add a structured per-paper LLM summary (problem, core insight, method, key results, positioning) | | `search_by_pdf_url` | Give it a public PDF link (e.g. an arXiv PDF) and get semantically similar papers from the index | | `search_research_topics` | A different corpus: open research problems mined from the limitations authors state in their own papers. Describe what you want to work on as prose — keywords match nothing, since each problem is stored as a synthetic abstract | | `list_indexes` | List the available paper indexes and their sizes | ## Setup ### Claude Code ```bash claude mcp add priorwork -- uvx priorwork-mcp ``` ### Cursor / other MCP clients Add to your MCP configuration: ```json { "mcpServers": { "priorwork": { "command": "uvx", "args": ["priorwork-mcp"] } } } ``` ## Caching The topics artifact is ~4.6 MB and changes only when the pipeline runs, so it is cached at `~/.cache/priorwork-mcp/topics.json` rather than fetched per search. Set `PRIORWORK_CACHE_DIR` to move it. Every search response names the published set it came from, so knowing when to refetch is one string comparison. That covers the change a client cannot infer for itself: a run that rewrites every title leaves every key intact, since a key derives from a topic's supporting papers rather than its text. A refetch that fails leaves the cached copy in place — stale topics beat no t…