Fieldry

Use case

Give Claude and Cursor grounded doc tools via MCP

Parse, tables and evidence as agent-callable tools — chunks with bbox for RAG, cited values for answers.

Get Started

The problem

Agents hallucinate about documents because they read dumped text with no structure and no source links.

  1. 01Flat text dumps. Whole PDFs pasted into context lose tables, layout and field boundaries — then the model guesses.
  2. 02No citations. Agent answers about documents can't point at sources, so every claim needs manual verification.
  3. 03Chunking without boxes. RAG chunks split mid-table and mid-sentence with no page coordinates to ground retrieval.
  4. 04Glue code tax. Every team hand-rolls parse→chunk→cite pipelines instead of calling one tool.

The solution

Fieldry's MCP server gives agents parsing, tables and evidence as tools — same cited output as the API.

  1. 01MCP-native tools. Claude, ChatGPT, Cursor and Agent2Agent call extract, parse and tables as tools with typed results.
  2. 02Layout-aware chunks. Parse returns chunks with bounding boxes — retrieval grounds to page coordinates, not guesses.
  3. 03Cited answers. Values arrive with confidence and evidence, so agents can quote sources inline.
  4. 04Same meter. Tool calls meter like API calls (4/page, +2 tables, −50% batch) with X-Credits-Used in results.

Agent tools

Search the tool calls.
See what each returns.

The MCP tools agents call — same shapes as the REST API.

MCP-native tools

Claude, ChatGPT, Cursor and Agent2Agent call extract, parse and tables as tools with typed results.

Layout-aware chunks

Parse returns chunks with bounding boxes — retrieval grounds to page coordinates, not guesses.

Cited answers

Values arrive with confidence and evidence, so agents can quote sources inline.

Command preview

Typed tool result

Agents pass a file and schema; the tool returns schema-valid JSON with evidence per value.

Anatomy of an agent query

A user asks Claude for last quarter's freight spend by carrier from a folder of carrier invoices.

Claude calls extract with the invoice schema, sums the cited rows, and answers with per-invoice sources attached.

  • InputFolder of carrier PDFs
  • OutputSummed answer + per-row sources
  • ReviewSources clickable, no blind trust

From question to cited answer

  1. Human reads the answer with sources attached — verification is one click.

    Analyst

  2. Agent aggregates cited rows; totals reconcile against the batch.

    Fieldry

  3. Agent calls extract via MCP; rows return cited to page and bbox.

    Fieldry

  4. Invoice folder sits unstructured — no rows, no sources, no sums.

    Document

4 milestones, newest first.

Key benefits

Agents that cite instead of invent.

Zero glue code

One MCP server replaces hand-rolled parse→chunk→cite pipelines.

Grounded RAG

Chunks carry page coordinates; retrieval hits land on real lines.

Works everywhere

Claude, ChatGPT, Cursor, n8n — same tools, same cited shapes.

Transparent meter

Tool calls report credits like API calls. No surprise bills from chatty agents.

Plug parsing into your agent

Connect the MCP server and ask about your own documents with sources attached.

Get started

Agent & RAG questions, answered

The questions teams ask most about agent & rag.

Talk to us

Any MCP client: Claude, ChatGPT, Cursor, plus Agent2Agent orchestrators. REST, SDKs and n8n cover non-MCP stacks with identical output shapes.

Values arrive with page, source text and bbox, so the agent can quote or link the exact line instead of asserting from memory.

Tool calls meter like API calls (4/page, +2 tables, −50% batch) with X-Credits-Used in results — set per-agent budgets from the headers.

Yes — the SDK waiter auto-polls, or subscribe to the signed webhook. Same payload either way.

Parse your first document today

One API for extraction, evidence and automation.

Request a DemoSee pricing