doc.cheap MCP server – passport, ID card and MRZ OCR for MCP clients

Give your assistant a passport, national ID card or driver's licence and get the printed fields back as structured JSON – $0.01 per recognised document, and free to try before you register. Without a key, the public sandbox key is used. It gives 10 free recognised documents per address in all, and at most 10 requests per address an hour, whatever their answer. Registering gives 100 free documents every month. Use test images only, never a real document.

An MCP server over stdio, for Claude Desktop, Claude Code, Cursor, VS Code, Gemini CLI, Windsurf, Kiro and any other MCP client – and the same server hosted at https://mcp.doc.cheap/mcp for clients that connect to a URL instead. It is a thin client of the public doc.cheap HTTP API and a copy of the documentation: it holds no data of its own.

npx -y @doc-cheap/mcp

Hosted – nothing to install

https://mcp.doc.cheap/mcp serves the same six tools over Streamable HTTP. No login: send your key as X-Doc-Cheap-Api-Key: sk_live_your_key or Authorization: Bearer sk_live_your_key – the named header wins when both are sent, and an Authorization that is not a doc.cheap key is ignored rather than passed on. Without a key, the public sandbox key is used. It gives 10 free recognised documents per address in all, and at most 10 requests per address an hour, whatever their answer. Registering gives 100 free documents every month. Use test images only, never a real document. The hosted server cannot read files on your machine, so scan_document takes the image as image_base64 or image_url; image_path is for the local server only.

Claude Code:

claude mcp add --transport http doc-cheap https://mcp.doc.cheap/mcp --header "Authorization: Bearer sk_live_your_key"

Claude Desktop and claude.ai: Settings → Connectors → Add custom connector, URL https://mcp.doc.cheap/mcp (no key: the sandbox key is used).

Cursor, ~/.cursor/mcp.json:

{
  "mcpServers": {
    "doc-cheap": {
      "url": "https://mcp.doc.cheap/mcp",
      "headers": { "Authorization": "Bearer sk_live_your_key" }
    }
  }
}

VS Code, .vscode/mcp.json (the key is asked for once and stored by VS Code):

{
  "servers": {
    "doc-cheap": {
      "type": "http",
      "url": "https://mcp.doc.cheap/mcp",
      "headers": { "Authorization": "Bearer ${input:doc-cheap-key}" }
    }
  },
  "inputs": [
    {
      "type": "promptString",
      "id": "doc-cheap-key",
      "description": "doc.cheap API key",
      "password": true
    }
  ]
}

Related MCP server: document-intelligence-server

Install locally

Set DOC_CHEAP_API_KEY to your key. Without a key, the public sandbox key is used. It gives 10 free recognised documents per address in all, and at most 10 requests per address an hour, whatever their answer. Registering gives 100 free documents every month. Use test images only, never a real document. The sandbox key has no balance.

Claude Desktop – one-click extension

Download doc-cheap-<version>.mcpb from the latest release and open it, or drag it into the Claude Desktop window. The install screen asks for two optional settings: your API key (stored as a secret; leave it empty for the sandbox key) and the one folder image_path may read images from (leave it empty and no local file is read). Node.js ships with Claude Desktop, so nothing else needs installing.

Claude Desktop – by hand

claude_desktop_config.json:

{
  "mcpServers": {
    "doc-cheap": {
      "command": "npx",
      "args": ["-y", "@doc-cheap/mcp"],
      "env": { "DOC_CHEAP_API_KEY": "sk_live_your_key" }
    }
  }
}

Claude Code

claude mcp add-json doc-cheap '{"command":"npx","args":["-y","@doc-cheap/mcp"],"env":{"DOC_CHEAP_API_KEY":"sk_live_your_key"}}'

Cursor

~/.cursor/mcp.json (or .cursor/mcp.json in a project):

{
  "mcpServers": {
    "doc-cheap": {
      "command": "npx",
      "args": ["-y", "@doc-cheap/mcp"],
      "env": { "DOC_CHEAP_API_KEY": "sk_live_your_key" }
    }
  }
}

VS Code

code --add-mcp '{"name":"doc-cheap","command":"npx","args":["-y","@doc-cheap/mcp"]}'

Or .vscode/mcp.json, which nests servers under servers rather than mcpServers:

{
  "servers": {
    "doc-cheap": {
      "type": "stdio",
      "command": "npx",
      "args": ["-y", "@doc-cheap/mcp"],
      "env": { "DOC_CHEAP_API_KEY": "sk_live_your_key" }
    }
  }
}

Gemini CLI

~/.gemini/settings.json:

{
  "mcpServers": {
    "doc-cheap": {
      "command": "npx",
      "args": ["-y", "@doc-cheap/mcp"],
      "env": { "DOC_CHEAP_API_KEY": "$DOC_CHEAP_API_KEY" }
    }
  }
}

The repository also carries gemini-extension.json, so it installs as a Gemini CLI extension without writing settings by hand.

Windsurf

~/.codeium/windsurf/mcp_config.json:

{
  "mcpServers": {
    "doc-cheap": {
      "command": "npx",
      "args": ["-y", "@doc-cheap/mcp"],
      "env": { "DOC_CHEAP_API_KEY": "${DOC_CHEAP_API_KEY}" }
    }
  }
}

Kiro

.kiro/settings/mcp.json in the workspace, or ~/.kiro/settings/mcp.json:

{
  "mcpServers": {
    "doc-cheap": {
      "command": "npx",
      "args": ["-y", "@doc-cheap/mcp"],
      "env": { "DOC_CHEAP_API_KEY": "${DOC_CHEAP_API_KEY}" },
      "disabled": false,
      "autoApprove": ["check_balance", "search_docs"]
    }
  }
}

Kiro also installs from a one-click link, which writes that block for you – it asks for confirmation first and shows the command and argument list it is about to add:

https://kiro.dev/launch/mcp/add?name=doc-cheap&config=%7B%22command%22%3A%22npx%22%2C%22args%22%3A%5B%22-y%22%2C%22%40doc-cheap%2Fmcp%22%5D%2C%22disabled%22%3Afalse%7D

(config is the URL-encoded JSON of {"command":"npx","args":["-y","@doc-cheap/mcp"],"disabled":false}.)

Claude Code plugin

The repository carries .claude-plugin/plugin.json, so it installs as a Claude Code plugin rather than as a hand-written server entry.

Every client starts the server as a process, so an edited configuration takes effect on the client's next launch.

Tools

scan_document

Recognise a passport, national ID card or driver's licence and return what is printed on it. Give it the image as image_base64, image_path or image_url, plus the optional expect_country, return_portrait, retain_hours, reference and idempotency_key.

{
  "image_base64": "/9j/4AAQSkZJRgABAQ…",
  "expect_country": "GRC",
  "idempotency_key": "order-4711-front"
}

It answers with the whole Scan as structured JSON – meta (id, status, billed, confidence, timing), document (kind, issuing country, number, series, date of issue, date of expiry, whether it has expired and how many days are left), holder (given names, surname, date of birth, sex, nationality; each country code comes with a *_detail object holding the code as the document writes it, its ISO 3166-1 alpha-3 and alpha-2 and its category), fields (every field read off the printed page, each with its own confidence), mrz (whether the machine-readable zone checks out, why not when it does not, and its lines exactly as read), images, quality and authenticity – and a one-line summary of the same result:

Scan 01a0af18-cd8d-7a61-9f2d-4c7b8e105da3 · recognized · passport (GRC) · PARADEIGMA ELENI SOFIA · billed · 684 ms

One recognised document costs one credit, $0.01. An unreadable image, an empty frame or an unsupported type costs nothing, and meta.billed says which happened. Sending the same idempotency_key again returns the first result rather than recognising and charging a second time.

check_balance

No arguments. Returns the balance – this month's free credits (100 every month, drawn first) and the paid credits, each on its own – the credits spent and this period's scan counters by status. With the public sandbox key there is no account behind the call: the balance comes back as null, and the first line says so and how to get a key, instead of reporting zeros that read like a balance.

Balance: 1840 credits · 42 free credits left this month · 1798 paid credits · 63 scans this period (58 billed, 58 credits spent).

search_docs

{ "query": "mrz check digit", "limit": 5 }

Full-text search over the documentation – endpoints, response fields, error codes, MRZ rules, retention, pricing – returning the matching sections with titles, snippets and links. It reads a copy shipped inside this package, so it makes no network call.

Stored scans: list_scans, get_scan, delete_scan

A result is stored only when the scan was made with a live key under a non-zero retention window – retain_hours on the call, or the account's history-retention setting (one year by default) when the call named none – and only until that window ends. A key reaches its own account's scans and no other's. Under a sandbox key nothing can be read back: an account's own sandbox key stores nothing, and the public key sk_sandbox_public keeps every scan for one year, with a reduced thumbnail and the address it came from, where that key cannot reach it. So the list is empty and a lookup or a deletion finds nothing; with the public sandbox key the server says so without calling the API. None of the three charges a credit.

list_scans returns the stored scans, newest first, one page at a time:

{ "limit": 20 }

Each row carries id, status, billed, duration_ms, reference and created_at, and the page carries next_cursor – pass it back as cursor for the next page; it is null on the last one. limit is 1 to 100 and defaults to 20. A row holds no extracted data.

get_scan takes a scan_id – meta.id of a scan_document result, or id of a row – and returns the full Scan as it was first returned, except that the image crops are never stored (every images slot is null) and quality reads not_checked. It never recognises the image again.

delete_scan takes a scan_id and permanently deletes that stored scan – the result, its history row and its thumbnail. It cannot be undone: the scan can no longer be listed, fetched or replayed through its idempotency_key. The credit it drew is not refunded. It answers { "id": "…", "deleted": true }.

An id that is unknown, belongs to another account, was made with retain_hours: 0 or has passed its window is not_found for both tools; the cases are not told apart.

Every tool declares an output schema, and every successful call returns structuredContent that matches it, beside the text blocks: the API's Scan for scan_document and get_scan, its Usage for check_balance, a page of { "scans": [...], "next_cursor": … } for list_scans, { "id", "deleted" } for delete_scan, and { "results": [...] } for search_docs. A failed call is an error block with no structured content.

Resources

Every page of the documentation this package ships is a read-only resource, text/markdown, with its title and description:

doccheap://docs/reference/fields
doccheap://docs/reference/errors
doccheap://docs/errors/insufficient_credits

resources/list lists them all, and the template doccheap://docs/{+slug} looks one up by its path. Reading one makes no network call.

Prompts

Configuration

Image sources

The server runs on your machine, with your files and your network, and the arguments to scan_document are chosen by a model. So the two sources that are not the image itself are fenced in:

• image_path is disabled until you set DOC_CHEAP_IMAGE_ROOT to a directory of your choosing. With it set, only files inside that directory can be read: both the directory and the requested file are resolved to their real locations first, so .. segments and symlinks pointing out of the directory are refused rather than followed. A relative image_path is taken from that directory. Without the variable the tool answers with an error telling the agent to set it or to send image_base64.

• image_url must be https: and must resolve to a public internet address. Loopback, private, link-local (including the cloud metadata address), carrier-grade NAT, multicast, reserved and IPv6 unique-local and link-local addresses are refused, as are the IPv4-mapped IPv6 spellings of them. Redirects are followed by hand, at most three hops, and every hop is re-checked, so a public URL cannot hand off to a private one. The body is capped at 25 MB – the API refuses more anyway.

• image_base64 has no such constraints: the caller already holds the bytes. It is the fallback every refusal above points at.

A guard refusal is a normal tool error with a readable message, so the agent can tell you what to change.

What this server sends about itself

When it calls the doc.cheap API it identifies itself in the request's User-Agent, the way any HTTP client does:

doc-cheap-mcp/0.3.11 (claude-code/1.4.2)

The first half is this package and its version. The second half is the name and version your MCP client reports over the protocol – the editor or assistant you launched it from – normalised to a short label, plus the same label on a baggage header. It is used for one thing: counting how much this server is used and from which applications, so that the work goes where people actually are. It is never used to change what the server does, and nothing else about you, your prompts, your files or your images travels with it.

Switching it off: set DO_NOT_TRACK=1 in the server's environment. The request then carries doc-cheap-mcp/0.3.11 and nothing more – no client name, no client version, no baggage header – and everything else works identically.

Your API key already identifies your account to the API; that is what a key is for, and it is unaffected by the setting above.

Privacy Policy

The full policy is https://doc.cheap/privacy. What it says about this server:

• What is collected. The image you ask it to read, sent to the doc.cheap API (https://api.doc.cheap) – nowhere else. Its calls also name this package and version, and the client you run it in (see the section above; DO_NOT_TRACK=1 removes the client). Your key identifies your account to the API.

• How it is used and stored. The image is read and never stored: it lives in memory for the length of the request. The result – the fields read off the document – is kept for the window the call asked for in retain_hours (0 stores nothing) or, when it asked for none, for the account's history-retention setting, which defaults to one year; expiry deletes it, and delete_scan deletes one result sooner.

• Who else sees it. Nobody the policy does not name: the hosting provider and the network provider that carry the traffic. Nothing is sold or shared for advertising. Nothing this server does is reported anywhere unless you set DOC_CHEAP_SENTRY_DSN yourself.

• On your machine. The server reads no file unless you name a folder for image_path, and then only images inside it.

• Contact. admin@doc.cheap.

The hosted server keeps nothing either. Its log records which method and which tool a request called, how long it took and whether a key of your own was used – never the image, the result, the key or your address.

Run it from source

{
  "mcpServers": {
    "doc-cheap": {
      "command": "node",
      "args": ["/absolute/path/to/the/checkout/apps/mcp/src/index.ts"],
      "env": { "DOC_CHEAP_API_BASE": "http://127.0.0.1:3000" }
    }
  }
}

pnpm --filter @doc-cheap/mcp build bundles the server to build/index.js with a shebang and copies the documentation content next to it, so the bin (doc-cheap-mcp) runs standalone.

Licence

MIT – see LICENSE. The monorepo this server is developed in is UNLICENSED; this package alone is published, and it is published under MIT.

Links

• Guide: https://doc.cheap/docs/guides/use-the-mcp-server

• Documentation: https://doc.cheap/docs

• Get an API key: https://doc.cheap/register

• Package: https://www.npmjs.com/package/@doc-cheap/mcp

• Source and issues: https://gitlab.com/doccheap/ocr-mcp

• doc.cheap: https://doc.cheap

In the MCP registry this server is cheap.doc/mcp.