paperqa-mcp-server
Provides an MCP server that wraps PaperQA2 to index a local PDF library and answer questions by reading/searching/synthesizing across those papers. Indexing uses OpenAI embeddings and (optionally) other LLMs configured via environment variables; the tool is intended for use inside Claude Desktop via MCP.
Score Breakdown
⚙ Agent Friendliness
🔒 Security
Security posture is inferred from documentation: communications to model providers occur over TLS in typical client libraries (not explicitly stated). Auth is via provider API keys in environment variables (stronger than no auth, but no scoped/least-privilege model is described). The README instructs users to place OPENAI_API_KEY directly into Claude Desktop config env; while convenient, this increases risk of accidental exposure if config files/logging are mishandled. The tool processes local PDFs and sends chunks to external embeddings/LLM providers during indexing and query synthesis, so data exposure depends on provider terms and the user's environment; no explicit redaction/PII controls are described.
⚡ Reliability
Best When
You have a stable local PDF library (e.g., Zotero storage), can run/refresh the index periodically, and are comfortable sending extracted text/chunks to external model providers for embeddings and synthesis.
Avoid When
You cannot provide OpenAI credentials, cannot run a local indexing step, or require strict on-prem data residency where OpenAI calls are not allowed.
Use Cases
- • RAG-style Q&A over a local collection of scientific PDFs
- • Cross-paper synthesis (compare methods, results, assumptions)
- • Building evidence-backed summaries with citations pointing to local Zotero storage keys
- • Assisting research workflows by combining paper text with Zotero metadata via zotero-mcp
Not For
- • Unindexed or frequently changing PDF libraries without running the index step (server refuses to answer if too many papers are unindexed)
- • Use cases requiring a public hosted API, webhooks, or direct REST/GraphQL access
- • Handling highly sensitive regulated workloads without ensuring data processing policies for external LLM/embedding providers
Interface
Authentication
Authentication appears to be done via environment variables for model providers (OpenAI; optionally Anthropic). No user-level OAuth/session flow or fine-grained scopes are described.
Pricing
Cost depends on the number of papers/chunks and the configured embedding/LLM models.
Agent Metadata
Known Gotchas
- ⚠ Claude Desktop has a short startup timeout; initial dependency downloads can cause 'server disconnected'—pre-warm by running uvx paperqa-mcp-server once.
- ⚠ If too many PDFs are unindexed, the server refuses to answer and instructs running the index command first.
- ⚠ Claude Desktop config JSON must be valid; missing commas can prevent the MCP server from loading.
- ⚠ The index hash is sensitive to settings; building the index with the server's own index command is required to ensure the server finds it.
- ⚠ Indexing may hit external rate limits; the documented recovery is to re-run the same index command.
Alternatives
Full Evaluation Report
Comprehensive deep-dive: security analysis, reliability audit, agent experience review, cost modeling, competitive positioning, and improvement roadmap for paperqa-mcp-server.
AI-powered analysis · PDF + markdown · Delivered within 30 minutes
Package Brief
Quick verdict, integration guide, cost projections, gotchas with workarounds, and alternatives comparison.
Delivered within 10 minutes
Score Monitoring
Get alerted when this package's AF, security, or reliability scores change significantly. Stay ahead of regressions.
Continuous monitoring
Scores are editorial opinions as of 2026-04-04.