Nanonets MCP server (DocStrange): what it does, what it can read, how to connect

Nanonets documents one MCP server: DocStrange, a local, read-only developer tool with 11 PDF tools, run from a cloned repo. Nanonets has no hosted server. Verified 21 Oct 2026.

Official MCP server · read-only (local, developer setup)

Hosting: Local stdio only. The server lives in the open-source NanoNets/docstrange repo (mcp_server_module/server.py) and is not in the PyPI package
Auth: None for the server itself. server.py passes no key, but the DocStrange library automatically picks up a NANONETS_API_KEY environment variable or a cached docstrange login token, then sends documents to Nanonets' cloud under that account
Read / write: Read-only: parses PDFs you point it at and returns text, sections and chunks. No tool changes data in Nanonets or writes to disk, but by default each PDF is sent to Nanonets' cloud for processing
Works with: Claude Desktop (documented). Other MCP clients that run local stdio servers, such as Claude Code. Not ChatGPT
Plan requirement: Not documented for the MCP server. The DocStrange README says signed-in or API-key cloud use is free up to 10,000 documents a month
Last changed: 2025-08-07: last commit to the MCP server module (added 6 Aug 2025)
Verification: Verified 21 Oct 2026 by theFinanceStack · confidence: medium

What can an agent do with Nanonets through MCP?

An agent can read PDFs on your machine through 11 DocStrange tools. It cannot reach Nanonets' hosted AP or workflow platform.

This is a developer tool. The server wraps DocStrange, Nanonets' open-source document extraction library (MIT), so Claude can turn a PDF into markdown and then search and read it by section.

Tools

The source defines 11 tools.

Group	Tools
Basic (7)	Parse a PDF, list headers, search, get one section, section summaries, full content, list cached documents
Long documents (4)	Token stats, section tree, split into chunks of up to 4,000 tokens by default, read one chunk

Good to know

Parsed documents are kept in an in-memory cache.
No tool creates, posts or approves anything, and no tool moves money.
The Nanonets homepage says agents can connect "through an MCP server". The Nanonets docs have no MCP page, so that claim about the hosted platform is not documented.

Can

Convert a local PDF, such as an invoice, contract or bank statement, to markdown (local)
List a document's headings, search it for a phrase and return one section by title (local)
Summarise every section, or return the full parsed text (local)
Report token counts and split a long document into chunks for step-by-step reading (local)

Cannot

Read or change data in the Nanonets platform (workflows, approvals, exports); the Nanonets docs describe no MCP server for the platform
Write, post or approve anything: all 11 tools are read-only per the server source. Files still leave your machine: by default they go to Nanonets' cloud (see cautions)
Run as a hosted connector: per the README it is for local development and is not shipped in the PyPI package

What access does it get?

The server reads any file path the agent passes it, with your user account's file permissions. There are no OAuth scopes.

Where	Scopes and permissions (as documented)	Admin requirement
DocStrange MCP server (local stdio)	No scopes. Not documented by the vendor beyond tools that take a `file_path`. Per `server.py`, the server resolves any path and checks only that the file exists, with no folder allowlist, so it can open any file your user account can read. The parsed text goes to your AI client and, in the default cloud mode, the file goes to Nanonets. Run it under a user or folder that holds only the documents meant for parsing.	Nanonets doesn't document one. Claude Team/Enterprise: an Owner may restrict which MCP servers Claude Code can run. The library uses a `NANONETS_API_KEY` or cached `docstrange login` token automatically if one is on the machine; a Nanonets API key comes from app.nanonets.com.

How do I connect it?

Clone the repo, install it in development mode, point Claude Desktop at server.py and restart. Plan on about four steps, plus Python 3.10 or later.

Nanonets doesn't publish a hosted server or Claude directory connector as of 21 Oct 2026. Claude Desktop uses the local config (see the Local tab).

What should a finance team watch for?

Assume documents go to Nanonets' cloud: the DocStrange README and source make cloud processing the default, even though the MCP README says local.

Files are sent to Nanonets' cloud by default, per the DocStrange README and source (server.py calls DocumentExtractor() with no arguments, and extractor.py sets cloud mode unless GPU mode is set). The MCP README's line that the server uses EasyOCR locally conflicts with the code. Local-only processing needs editing server.py to pass gpu=True (and a GPU), per the MCP README. If a NANONETS_API_KEY or cached docstrange login token exists on the machine, parsing runs under that account.
Without an API key or login, DocStrange cloud use is rate-limited; the README says "limited calls daily" and gives no exact number. Signed-in or API-key use is free up to 10,000 documents a month, per the README. That is the DocStrange free tier; the README doesn't relate it to Nanonets' platform pricing.
The MCP README shows the API key set in server.py as a string. Because the library reads NANONETS_API_KEY, the key can go in the environment or the client config's env block instead of a git checkout. docstrange login stores its token in a plain-text file at ~/.docstrange/credentials.json, which the MCP server uses automatically. Keys in config sit in plain text and often get synced or committed; don't commit them, and use your client's secret storage where it has one.
Parsed text from supplier invoices, emails or contracts is free text. An injected instruction could steer the agent's next actions in other connected tools, so keep human approval on anything that writes elsewhere.
Nanonets' DocStrange README says nothing on cloud retention or model training. Since cloud is the default path, check Nanonets' privacy policy and terms (linked in sources) and your AI provider's terms.
The server is a development setup. It has no versioned releases or PyPI package, and the MCP module has had no commits since August 2025.

How could finance teams and accounting firms use it?

Use it to read and question documents. It cannot change any ledger or Nanonets workflow.

Inside a company

Ask questions of a long supplier contract, such as payment terms or renewal dates, section by section.
Pull headline figures and dates from a PDF invoice or statement for a person to check and key in.
Summarise a lengthy audit or board pack before a review meeting.
Search a set of policy PDFs for a clause, such as an expense limit.

At an accounting firm

Read a new client's engagement documents or prior-year statements and list open questions.
Pull section summaries from a loan agreement to draft a covenant checklist for a reviewer.
Extract figures quoted in a PDF report so a reviewer can compare them with the client's records and sign off any differences.
Have a developer test the setup first, including where processing runs, before any client files go through it.

Keep a person checking anything taken from a parsed document before it changes the ledger or any system of record.

Want Nanonets connected to Claude for your team?

We configure scoped connections and a review step before anything changes your books, for a fixed fee.

Book a 30-minute setup call

Change log

Date	Change	Source
2025-08-07	"Updated docstrange library function": last commit to the MCP server module	github.com/NanoNets/docstrange/commits/main/mcp_server_modul
2025-08-06	MCP server added to the DocStrange repo	github.com/NanoNets/docstrange/commits/main/mcp_server_modul
2025-08	DocStrange README announces a 7B-parameter model upgrade and a local web interface	github.com/NanoNets/docstrange

FAQ

Is this official?

Yes. The server is in the NanoNets GitHub organisation's DocStrange repo (MIT). Nanonets' homepage also mentions an MCP server for its platform, but its docs don't describe one.

Is it free?

The code is open source. The DocStrange README says cloud processing is free up to 10,000 documents a month when signed in or using an API key, and rate-limited without one.

Does it change data in Nanonets?

No. None of the 11 tools change data in Nanonets or your ledger, per the server source. By default, though, parsing uses DocStrange cloud mode, which sends the document to Nanonets for processing.

Is my data used to train models?

The DocStrange README doesn't say. Check Nanonets' terms for cloud processing and your AI provider's terms, and confirm whether your setup processes files locally.

Can my accountant use it on my behalf?

Nanonets doesn't address this. The server runs on one machine and reads files there, so your accountant would install it and parse documents you share with them.

Something wrong or out of date? Tell us and we'll email you when it's fixed.

Sources

Verified by theFinanceStack, 21 Oct 2026. Vendor list on the CFO Software Market Map courtesy of Jason Staats; AI connectivity data by theFinanceStack. Not tax, compliance or financial advice.