scan-mcp
Enables scanner capture via SANE on Linux, providing tools for device discovery, scan jobs with ADF/duplex support, document batching, and multipage assembly through a local MCP server.
README
<p align="center"> <img src="docs/assets/icon.png" alt="scan-mcp logo" width="96"> </p>
<h1 align="center">scan-mcp</h1>
Minimal MCP server for scanner capture (ADF/duplex/page-size), batching, and multipage assembly.
Features
- Small, typed MCP server exposing tools for device discovery and scan jobs
- JSON Schema–validated inputs with deterministic, typed outputs
- Smart device selection (prefers ADF/duplex, avoids camera backends), robust defaults
- Local-first transports: stdio by default to keep everything on-device, optional HTTP for your own network deployments
Note: This package targets Node 22 and Linux SANE backends (scanimage).
Quick Start (local stdio, default)
Add a server entry to your MCP client configuration:
{
"mcpServers": {
"scan": {
"command": "npx",
"args": [
"-y",
"scan-mcp"
],
"env": {
"INBOX_DIR": "~/Documents/scanned_documents/inbox"
}
}
}
}
- This invocation runs over stdio for a privacy-first, single-machine setup.
- Call
start_scan_jobwithout adevice_idto auto-select a scanner and begin scanning. - Artifacts are written under
INBOX_DIRper job:job-*/page_*.tiff,doc_*.tiff,manifest.json,events.jsonl. Whencrop_carrier_sheetsis set and a carrier sheet is detected, apage_*.cropped.tiffderivative is also written per affected page.
Streamable HTTP transport
Prefer to attach the scanner to another machine on your network? scan-mcp also supports the
streamable HTTP transport:
scan-mcp --http
- Default port is
3001; setMCP_HTTP_PORTto override (for exampleMCP_HTTP_PORT=3333 scan-mcp --http). - Binds all interfaces (
::) by default; setMCP_HTTP_HOSTto restrict (for exampleMCP_HTTP_HOST=127.0.0.1when a reverse proxy fronts the server). - HTTP responses use server-sent events (SSE) for streaming tool output; clients such as Claude Desktop and Windsurf support this transport.
- There is currently no authentication; this is intended for internal LAN networking
Install
- Run with npx:
npx scan-mcp(recommended)- The CLI runs a quick preflight check for Node 22+ and required scanner/image tools and prints installation hints if anything is missing.
- See recommended server config above
- Use
npx scan-mcp --httpto launch the streamable HTTP transport when running on another machine. - CLI help:
scan-mcp --help - From source (for development):
npm installnpm run build
- For Cline setup, and other automated agentic installation, see llms-install.md
System Requirements
- Linux with SANE utilities:
scanimage(and optionallyscanadf) - TIFF tools:
tiffcp(preferred) or ImageMagickconvert
Environment Variables
SCAN_MOCK(default:false): mock SANE calls and generate fake TIFFs for testing.INBOX_DIR(default:scanned_documents/inbox): base directory for job runs and artifacts.SCANIMAGE_BIN/SCANADF_BIN(defaults:scanimage/scanadf): override binary paths.TIFFCP_BIN/IM_CONVERT_BIN(defaults:tiffcp/convert): multipage assembly tools.SCAN_EXCLUDE_BACKENDS(CSV): backends to exclude (e.g.,v4l).SCAN_PREFER_BACKENDS(CSV): preferred backends (e.g.,epjitsu,epson2).PERSIST_LAST_USED_DEVICE(default:true): persist and lightly prefer last used device.MCP_HTTP_PORT(default:3001): TCP port for the HTTP transport.
API
Tools
-
list_devices
- Discover connected scanners with backend details.
- Inputs: none.
-
get_device_options
- Get SANE options for a specific device.
- Inputs:
device_id(string): Target device identifier.
-
start_scan_job
- Begin a scanning job; omitting
device_idtriggers auto-selection and default options. - Inputs (all optional unless noted):
device_id(string)resolution_dpi(integer, 50–1200)color_mode(Color|Gray|Lineart): color_mode defaults to Lineart (document-first); at >= 600dpi it defaults to Color, since high-dpi capture usually means artwork/photos where 1-bit destroys information. Pass color_mode explicitly to override either default; high dpi is the only signal used.source(Flatbed|ADF|ADF Duplex)duplex(boolean)page_size(Letter|A4|Legal|Custom)custom_size_mm{width,height}doc_break_policy{type,blank_threshold,page_count,timer_ms,barcode_values}output_format(string, defaulttiff)tmp_dir(string)crop_carrier_sheets(boolean, defaultfalse): detect carrier-sheet leading-edge band and write cropped page derivatives; raw pages are kept
- Begin a scanning job; omitting
-
get_job_status
- Inspect job state and artifact counts.
- Inputs:
job_id(string)
-
cancel_job
- Request job cancellation; best effort during scan loops.
- Inputs:
job_id(string)
-
list_jobs
- List recent jobs from the inbox directory.
- Inputs (optional):
limit(integer, max 100)state(running|completed|cancelled|error|unknown)
-
get_manifest
- Fetch a job's
manifest.json. - Inputs:
job_id(string)
- Fetch a job's
-
get_events
- Retrieve a job's
events.jsonllog. - Inputs:
job_id(string)
- Retrieve a job's
See JSON Schemas in schemas/ for input shapes. Tests assert against these contracts.
How Selection and Defaults Work
Defaults aim for 300dpi, reasonable color mode, and ADF/duplex when available. Full details on scoring and fallbacks live in docs:
- Selection and defaults:
docs/SELECTION.md
Project Layout
src/mcp.ts— MCP server entry and tool registrationsrc/services/*— hardware interface and job orchestrationschemas/— JSON Schemas used for validation and testsdocs/— architecture, conventions, and deep dives
Development
npm run dev(stdio MCP server),npm run dev:http(HTTP transport)make verifyruns lint, typecheck, and tests- Conventions:
docs/CONVENTIONS.mdand architecture indocs/BLUEPRINT.md
Roadmap
Tracking ideas and future improvements are documented in docs/ROADMAP.md.
推荐服务器
Baidu Map
百度地图核心API现已全面兼容MCP协议,是国内首家兼容MCP协议的地图服务商。
Playwright MCP Server
一个模型上下文协议服务器,它使大型语言模型能够通过结构化的可访问性快照与网页进行交互,而无需视觉模型或屏幕截图。
Magic Component Platform (MCP)
一个由人工智能驱动的工具,可以从自然语言描述生成现代化的用户界面组件,并与流行的集成开发环境(IDE)集成,从而简化用户界面开发流程。
Audiense Insights MCP Server
通过模型上下文协议启用与 Audiense Insights 账户的交互,从而促进营销洞察和受众数据的提取和分析,包括人口统计信息、行为和影响者互动。
VeyraX
一个单一的 MCP 工具,连接你所有喜爱的工具:Gmail、日历以及其他 40 多个工具。
graphlit-mcp-server
模型上下文协议 (MCP) 服务器实现了 MCP 客户端与 Graphlit 服务之间的集成。 除了网络爬取之外,还可以将任何内容(从 Slack 到 Gmail 再到播客订阅源)导入到 Graphlit 项目中,然后从 MCP 客户端检索相关内容。
Kagi MCP Server
一个 MCP 服务器,集成了 Kagi 搜索功能和 Claude AI,使 Claude 能够在回答需要最新信息的问题时执行实时网络搜索。
e2b-mcp-server
使用 MCP 通过 e2b 运行代码。
Neon MCP Server
用于与 Neon 管理 API 和数据库交互的 MCP 服务器
Exa MCP Server
模型上下文协议(MCP)服务器允许像 Claude 这样的 AI 助手使用 Exa AI 搜索 API 进行网络搜索。这种设置允许 AI 模型以安全和受控的方式获取实时的网络信息。