Research MCP
Finvaulta's MCP server connects your chatbot to 30+ banks' sell-side research — with hierarchy-aware chunks, time decay, and cross-firm consensus. Ask global questions. Cite firm and date on every answer.
Example prompt
“Compare how major banks are positioning on duration into Q2. Weight recent notes higher. Show me the key table from each and run the implied spread change if 10y yields move 50bps.”
Your client orchestrates macro search, consensus comparison, and table math — each backed by extracted PDF structure, not scraped web text.
MCP tools
The server exposes persona-driven retrieval. Your LLM navigates hierarchies, compares institutions, and runs math — without seeing every chunk in the corpus.
Search across macro sections with exponential time decay. A match from this morning ranks above a stronger semantic hit from last year.
Surface the top arguments from JPM, Goldman, and others on the same asset — side by side, with firm and date on every claim.
Discover table chunk IDs by document before running calculations. Macro and consensus responses also include related tables from matching reports.
Run sandboxed pandas on a table chunk by ID. Use list_table_chunks to find the right chunk_id first.
Why it beats copy-paste RAG
Claude Desktop, Cursor, and any MCP-compatible client can call Finvaulta as a tool. No copy-paste from PDFs into the prompt window.
Scores decay with age. Hawkish pivots from 2024 do not drown out this week's desk notes when you ask about rates.
OCR tables are stored as dedicated chunks with titles and page refs. List them by document, then run math on the chunk you need.
The MCP endpoint authenticates against your Finvaulta account. Trial and active Pro plans get access; free accounts do not.
Under the hood
Documents pass through the same OCR and LLM extraction pipeline as the web app. A secondary indexing stage then builds hierarchy-tagged text chunks, separate table chunks, and embeddings — without blocking the primary document pipeline.
Retrieval combines pgvector similarity with PostgreSQL full-text search and entity tags from normalized extraction (tickers, themes, institutions). A SQL decay function down-weights stale matches at query time. Institution–security–theme edges are recorded for future graph-aware retrieval.
Get started
Active or trialing Pro unlocks MCP credentials in Settings.
Paste the MCP config (stdio or HTTP) into Claude Desktop, Cursor, or your agent stack.
Your model calls tools — macro themes, consensus comparison, table math — grounded in 30+ banks' research.
MCP access is included with Pro. Start a 14-day trial — card required.