Ask an AI agent to read an invoice and most document MCP servers will hand the whole thing back as Markdown. Nicely formatted. Still not data. The agent now has to hunt for the total, guess the date format and hope the line items survived the conversion. A document processing MCP server works the other way round: the agent sends the document, and gets back named fields it can act on. Vendor, invoice number, due date, total, line items, as JSON.
This guide covers the five MCP servers that do that: Parseur, Docsumo, Lido, Nutrient and Nanonets.
Key Takeaways
- A document processing MCP server returns named fields and line items as JSON. A converter such as MarkItDown returns Markdown the agent still has to read.
- MCP changes how an agent reaches the extraction engine, not how accurate that engine is. Judge the engine, then the tools it exposes.
- Parseur, Docsumo, Nutrient and Nanonets run hosted servers. Lido runs locally with npx, so every machine needs Node and its own API key.
- Only Parseur's MCP documentation covers email intake. The others take files, URLs or pasted content, and most documents still arrive as email attachments.
- Free tiers differ in kind: Parseur gives 20 pages a month, Lido 50 pages that never expire, Docsumo a 14-day trial of 1,000 pages. Nutrient and Nanonets do not publish theirs for MCP.
What is a document processing MCP server?
A document processing MCP server is a service that exposes document extraction tools to an AI agent through the Model Context Protocol. The agent connects once, discovers the tools on offer (upload a document, extract its fields, check whether processing finished, download the results), and calls them as it needs them. No custom integration code, no reading API docs on the agent's behalf.
MCP itself is the open standard Anthropic published in late 2024 so that AI clients such as Claude, ChatGPT, Cursor and VS Code could talk to outside tools in one common language. Any MCP client can use any MCP server. That is the whole point, and the reason a document processing vendor now ships one instead of a plugin per chat app.
Two things worth being clear about before the comparison. First, MCP changes how the agent reaches the extraction engine, not how good that engine is. A server wrapped around a weak OCR model returns weak results faster. Second, the tools a server exposes decide what the agent can do. A server with only an "extract" tool lets the agent read a document. A server with upload, extract, status, download and export tools lets the agent run the whole workflow, from a PDF arriving to a row landing in a spreadsheet.
How we evaluated these MCP servers
Each server was judged on what its own MCP documentation and tool reference say it can do. Seven criteria, in the order they tend to matter once the demo is over:
- Tool coverage. Can the agent upload a document, extract its fields, check status, download the results and push them somewhere? A server that only extracts leaves the rest of the workflow to you.
- Template-free extraction. Does the agent describe the fields it wants in plain words and get them back on any layout, or does someone first have to configure a document type per vendor?
- Input channels. Local file, public URL, email. Most documents a business processes arrive as email attachments, so a server that cannot receive email pushes that step back onto the human.
- Hosting and sign-in. A hosted server with OAuth sign-in means nothing to install and no key to paste. A local server run with npx means a Node process on every machine that uses it, and an API key to look after.
- Setup effort. Measured in steps from "I have an account" to "my agent extracted a field".
- Maintenance. Who runs the server, which MCP protocol versions it supports, and how recently it changed.
- Pricing. Whether MCP use costs extra, and what the free tier lets you try before paying.
Comparison table
| Server | Output | Hosting and sign-in | Email intake | Human review | Free tier | Best for |
|---|---|---|---|---|---|---|
| Parseur | JSON fields and line items, Excel and CSV export, webhooks | Hosted, OAuth 2.1 with dynamic client registration, nothing to install | Yes, every mailbox has an email address | Optional, in the Parseur app | 20 pages a month | Agents that run the whole intake-to-export loop |
| Docsumo | JSON, CSV and XML, plus searchable PDFs and OCR text | Hosted, API key as bearer token | Not documented for MCP | Yes, approval and validation tools, review links per document | 14-day trial, 1,000 pages | Finance teams with case and approval workflows |
| Lido | Columns you name, written to CSV or JSON | Local, run with npx, API key pasted once through a browser prompt | Web app only, not through MCP | Web app only, not through MCP | 50 pages, no expiry, then from $29 a month | Developers prototyping on a folder of local files |
| Nutrient | Fields with citations and confidence, Markdown, tables | Hosted with browser sign-in, or self-hosted Document Engine package | No | Not documented | Free plan per product, allowance not published | Teams that also need OCR, redaction and signing |
| Nanonets | JSON, Markdown, invoice and contract clause extraction | Hosted, OAuth 2.1 sign-in, DocStrange API key as fallback | Not documented for MCP | Not documented | Not published for MCP | Developers already on DocStrange |
Hosting, sign-in and free tier figures come from each vendor's MCP and pricing pages as of 29 October 2026. "Not documented" means the vendor's MCP pages do not mention it, not that the product lacks it elsewhere.
Best MCP servers for document processing
Each entry follows the same template so you can compare like with like: who it suits, what tools the agent gets, how setup goes, where it stops, and what it costs.
Parseur
Best for: agents that need to own the full loop, from a document arriving to a clean row landing in a spreadsheet or another app, without a person in between.
MCP tools available: Parseur's MCP server exposes the same operations as the Parseur app. The agent can create a mailbox and name the fields it wants, send a document as a file, a public URL or email text, create a one-time upload link for files on disk, wait for processing to finish, read the extracted fields and line items for one document, list documents, download everything a mailbox extracted as Excel or CSV, set up webhooks and export configurations, and tune the mailbox (AI instructions, splitting, sender filters, date and number formats). Tools only appear if the user's Parseur role allows them.
Setup: add https://api.parseur.com/mcp as a remote MCP server in any client that supports Streamable HTTP, sign in to Parseur when the browser opens, approve access. Parseur uses OAuth 2.1 with PKCE and dynamic client registration, so there is no client ID to request and no key to paste. Per-client walkthroughs for Claude, ChatGPT, Cursor, Claude Code and VS Code are on the agents page, and the generic steps on the MCP client page.
Limitations: how a file gets in depends on the client. In Claude you attach it to the chat and Claude uploads it. In Cursor, Claude Code and other agents with a shell, the agent uploads it from disk. In ChatGPT, which can do neither, you give it a public URL or paste the email text, or forward the document to the mailbox's email address. Reviewing and correcting an extraction happens in the Parseur app, not through an MCP tool.
Pricing: documents sent through MCP use the same page credits as any other channel. The free plan includes 20 pages a month, paid plans start from there. See Parseur pricing.
Docsumo
Best for: finance and operations teams that already run cases, validation checks and approval steps in Docsumo and want an agent to drive them.
MCP tools available: the broadest tool list of the five, with 32 documented in the main reference and more behind it. Documents: upload_document, upload_case_document, list_documents, get_document_data, download_document, doctype_retry_documents. Document types: read and edit the schema and fields, run a field test against a real document, pull accuracy analytics. Cases and workflows: create and update cases, run validations, rerun workflows, and update_workflow_approvals to approve or reject a run waiting at a human step. Plus lookup tables, webhook configuration and activity logs.
Setup: hosted at https://mcp.docsumo.com/mcp over Streamable HTTP. Generate an API key under Settings, pass it as a bearer token in the client's MCP config. Documented for Claude Code, Claude Desktop, Cursor, VS Code, Windsurf, ChatGPT and Gemini CLI.
Limitations: the API key grants full account access, including permanent deletion, and Docsumo says one key must not be shared across several end users. Local file paths are not accepted on the hosted server, documents go in as file content or by URL. Rate limit of 10 requests a second per account, and batch downloads cap at 25 documents per call. Fields are configured per document type before extraction, which is the opposite of describing them on the fly.
Pricing: uploads through MCP use the plan's document credits. A 14-day trial covers up to 1,000 pages. Paid plans are quoted on volume, with no published price.
Lido
Best for: a developer with a folder of PDFs on their laptop who wants an agent in Claude Code or Cursor to extract a few named columns this afternoon.
MCP tools available: four. authenticate opens a local browser page where you paste your Lido API key once. extract_file_data takes a PDF, image or scan and returns the columns you ask for. extraction_tips gives the agent advice on refining a shaky extraction. extractor_usage reports the pages left in your quota. Results are written to CSV or JSON. The documented examples cover batch extraction over a folder and matching purchase order lines against invoices.
Setup: one command in Claude Code: claude mcp add lido -- npx -y @lido-app/mcp-server. Other clients run the same npx command from their MCP config. Credentials are stored in a gitignored file inside the project directory.
Limitations: the server runs locally, so each machine that uses it needs Node and its own authenticated copy. Only local files are documented as input. Email ingestion, approval routing and team collaboration exist in the Lido web app but are not reachable through MCP. No status, list or download tools, since extraction returns in the same call.
Pricing: the MCP server is open source under the MIT license. Extraction consumes your Lido page quota. The free tier is 50 pages that do not expire, then the Standard plan is $29 a month for 100 pages and one seat.
Nutrient
Best for: teams whose documents need more than extraction, such as OCR on scans, conversion to PDF/A, redaction, watermarking or signing, and who want one server for all of it.
MCP tools available: Nutrient does not publish tool names on its MCP page, only capabilities. The hosted server extracts fields you describe in plain words and returns them with citations and a confidence figure, parses PDFs, scans and images into Markdown and tables, classifies documents by type, converts and OCRs files, merges and watermarks them, produces PDF/A and PDF/UA, signs and redacts, and reports remaining credits.
Setup: add https://mcp.nutrient.io/mcp as a remote HTTP server and sign in with a Nutrient account in the browser. No API key. Step-by-step guides exist for Claude, Claude Code and Codex. For on-premises processing there is a separate self-hosted package, @nutrient-sdk/document-engine-mcp-server, configured with your Document Engine URL and token. A third in-product server for Nutrient Workflow uses an API key and, at the time of writing, only works with desktop clients.
Limitations: no email intake. Human review is not described anywhere in the MCP documentation. The lack of a published tool list makes it hard to know before connecting what the agent can and cannot call. Three servers with three sign-in methods is a lot of surface for one vendor.
Pricing: Nutrient says every product has a free plan, but neither the MCP page nor the API pricing page states the allowance or the entry price. Credits are shared with your organisation's other Nutrient usage.
Nanonets
Best for: developers already using Nanonets DocStrange who want their coding agent to call the same extraction without writing HTTP code.
MCP tools available: the enterprise documentation lists 15 tools in three groups. Extraction: extract_sync for documents up to about five pages, get_result to poll an asynchronous job, list_results to see past jobs. Targeted extractors such as extract_to_markdown, extract_invoice and extract_contract_clauses. Once a document is extracted, the agent can query it again without reprocessing.
Setup: hosted at https://mcp.nanonets.com/mcp. In Claude Code: claude mcp add nanonets --transport http https://mcp.nanonets.com/mcp. Claude Desktop opens a browser for OAuth 2.1 sign-in on first use. A DOCSTRANGE_API_KEY environment variable works as a fallback.
Limitations: Nanonets maintains two sets of MCP docs (enterprise and DocStrange) that do not fully agree on authentication, and one mentions a localhost:8080/mcp setup that is not explained. The hosted MCP server dates from March 2026, so it is the youngest here. No email intake, status, review or export tools are documented for MCP.
Pricing: not published for the MCP server. Usage follows your DocStrange or Nanonets plan.
Last updated on
