Give Claude Your Documents via MCP

Agents hallucinate about documents for a boring reason

Paste a PDF into a chat and the model reads unstructured text with no field boundaries, no tables and no source links — then guesses with confidence. The fix isn't a smarter model. It's giving the agent tools: parse this file against a schema, return the rows, cite every value.

That is what the Fieldry MCP server does. One connection, and Claude, ChatGPT, Cursor or any Agent2Agent orchestrator can call extraction as a typed tool.

Step 1: Connect the server

Add the MCP server with your API key. From then on the agent sees four tools: extract (file + schema → cited JSON), parse (layout-aware chunks with bounding boxes), tables (rows, not markdown soup) and async batch with webhook or polling.

No glue code, no chunking pipeline to babysit. The tools return the same shapes as the REST API.

Step 2: Ask for something checkable

Try: "Total freight spend last quarter per carrier from this folder, with sources." The agent calls extract with an invoice schema, sums the cited rows, and answers with per-invoice sources attached:

€18.4k across 6 invoices — [p.2] [p.5] [p.9] → cited

Compare that with the pre-tool answer: a confident number from memory, zero sources, and your afternoon spent verifying.

Step 3: Ground your RAG the same way

parse returns chunks with bounding boxes, so retrieval grounds to page coordinates instead of guessing. Rebuild the index through the tool and your RAG hits land on real lines — quotes included.

What it costs

Tool calls meter exactly like API calls: 4 credits per page, +2 with tables, batch at half price, all reported in X-Credits-Used. Set per-agent budgets from the headers — chatty agents can't surprise you.