Docling MCP

io.github.docling-project由 GitHub 账号 docling-project 发布v3.5.0更新于 Nov 4, 2026

Convert PDFs and other documents to structured formats via Docling, for AI applications.

概览

AI 生成的概览

Docling MCP 可将 PDF 等文档转换为结构化的 DoclingDocument JSON 或 Markdown,并让助手生成新文档,支持本地或远程 Docling Serve 转换。

可以做什么
  • 将 PDF 等文档(本地文件或 URL)转换为结构化的 DoclingDocument JSON
  • 生成新文档:创建文档、添加标题、章节标题、段落与嵌套列表,并导出为 Markdown 或保存
  • 将转换或编写的文档导出为 Markdown,图片呈现方式可配置
  • 在内存中缓存已转换文档以加快后续操作
  • 转换时可选启用 OCR、表格结构识别和视觉模型图片描述
  • 可选的 RAG 工具组,支持 LlamaIndex、LlamaStack 及 Milvus 上传与检索
适用场景
适合需要让助手把 PDF 或扫描件读成结构化内容,用于摘要、信息抽取或 RAG 流程的场景,也适合让助手按 Docling 格式起草文档并导出为 Markdown。有 Docling Serve 实例时选远程模式,需要离线转换时选本地模式。
运行要求
以本地 Python 进程运行,通常通过 uvx 启动。远程模式(默认)需要 DOCLING_MCP_SERVICE_URL,通常还需 DOCLING_MCP_SERVICE_API_KEY 指向 Docling Serve 实例。本地模式需安装 docling-mcp[local] 附加组件,且 OCR 引擎须安装在转换后端。可选 RAG 工具组需要 LLM 端点或 LlamaStack 服务器。
安装前请注意
远程模式会把文档发送到所配置的 Docling Serve 端点,该端点可能是第三方托管服务,请确认文件去向。DOCLING_MCP_SERVICE_API_KEY 属于密钥,应通过客户端 env 配置传入,不要提交到代码库。启用图片描述会为每张图增加视觉模型推理,可能显著拖慢转换。本地模式会在本机下载并运行模型。

安装

在 SourceWeft 中

  1. 打开 控制台中的 Docling MCP,将其添加到工作区。
  2. 为需要使用其工具的对话启用该服务。

Desktop only,通过 STDIO。 STDIO 服务会启动本地进程,因此需要 SourceWeft 桌面宿主。

其他 MCP 客户端

参照 仓库 中的启动说明。

README

[Docling]

Docling MCP: making docling agentic

[CI] [PyPI version] [PyPI - Python Version] [uv] [Ruff] [Pydantic v2] [prek] [License MIT] [PyPI Downloads] [LF AI & Data] [MCP Registry] [docling-mcp MCP server] [MCP Toplist]

A document processing service using the Docling-MCP library and MCP (Model Context Protocol) for tool integration.

Overview

Docling MCP is a service that provides tools for document conversion, processing and generation. It uses the Docling library to convert PDF documents into structured formats and provides a caching mechanism to improve performance. The service exposes functionality through a set of tools that can be called by client applications.

Compatibility

docling-mcpMCP Python SDK
>=3.0.0mcp>=2.0.0
>=2.0.0,<3.0.0mcp>=1.9.4,<2.0.0

If your MCP client application has not yet migrated to MCP SDK v2, pin:

bash
pip install "docling-mcp<3.0.0"

See MIGRATION.md for the full migration guide.

Installation Options

Remote Mode (Recommended - Lightweight)

For users with access to Docling Serve API:

Getting Docling Serve: Visit docling-serve for installation guides. You can deploy it from published container images or look for managed Docling SaaS offerings.

bash
pip install docling-mcp

Then configure your environment:

bash
export DOCLING_MCP_SERVICE_URL=https://your-docling-service.example.comexport DOCLING_MCP_SERVICE_API_KEY=your-api-key-hereexport DOCLING_MCP_CONVERSION_MODE=remote

Local Mode (Full Features)

For users who need local conversion or don't have Docling Serve access:

bash
pip install docling-mcp[local]

Then configure your environment:

bash
export DOCLING_MCP_CONVERSION_MODE=local

Hybrid Mode (Best of Both)

Install with local support and enable automatic fallback:

bash
pip install docling-mcp[local]

Configure for remote with fallback:

bash
export DOCLING_MCP_SERVICE_URL=https://your-docling-service.example.comexport DOCLING_MCP_CONVERSION_MODE=remoteexport DOCLING_MCP_FALLBACK_TO_LOCAL=true

Features

  • Conversion tools:
  • Generation tools:
    • Document generation in DoclingDocument, which can be exported to multiple formats
  • Local document caching for improved performance
  • Support for local files and URLs as document sources
  • Memory management for handling large documents
  • Logging system for debugging and monitoring
  • RAG applications with Milvus upload and retrieval

Configuration

All settings use the DOCLING_MCP_ prefix and can be supplied as environment variables, in a .env file in the working directory, or via the env block of your MCP client config. Copy .env.example as a starting point.

Conversion mode

VariableDefaultDescription
DOCLING_MCP_CONVERSION_MODEremoteremote or local

Remote service (required when DOCLING_MCP_CONVERSION_MODE=remote)

VariableDefaultDescription
DOCLING_MCP_SERVICE_URL—URL of the Docling Serve instance
DOCLING_MCP_SERVICE_API_KEY—API key for the service
DOCLING_MCP_SERVICE_TIMEOUT300.0Timeout in seconds for a whole conversion job
DOCLING_MCP_SERVICE_MAX_RETRIES3Max retry attempts
DOCLING_MCP_FALLBACK_TO_LOCALfalseFall back to local if service is unreachable (requires docling-mcp[local])

Conversion pipeline (applies to both modes)

VariableDefaultDescription
DOCLING_MCP_KEEP_IMAGESfalseRetain page images in output
DOCLING_MCP_IMAGES_SCALE1.0Image scale factor (increase to avoid tensor padding errors)
DOCLING_MCP_DO_OCRtrueRun OCR pipeline
DOCLING_MCP_OCR_ENGINEautoOCR engine by its docling engine kind: auto, easyocr, tesseract (system binary), tesserocr (Python bindings), ocrmac, rapidocr. The engine must be installed on the conversion backend. In remote mode it is sent as the Docling Serve ocr_preset, so server-defined preset names also work there
DOCLING_MCP_OCR_LANG—Comma-separated OCR languages, passed to the engine unchanged. iso:en,iso:de works with any engine (docling ≥ 2.127, recommended with auto); bare codes are engine-specific (eng for Tesseract, en for EasyOCR or Docling Serve). Unset means the engine's defaults
DOCLING_MCP_DO_TABLE_STRUCTUREtrueDetect table structure
DOCLING_MCP_DO_PICTURE_DESCRIPTIONfalseRun a vision-language model to generate textual descriptions for figures and pictures
DOCLING_MCP_PICTURE_DESCRIPTION_PRESET—Preset ID for the picture-description VLM (e.g. smolvlm, granite_vision). When unset the backend default is used. Only applies when DOCLING_MCP_DO_PICTURE_DESCRIPTION=true

[!TIP] Enabling DOCLING_MCP_DO_PICTURE_DESCRIPTION=true adds a VLM inference step for every figure in the document. On resource-constrained hardware (e.g. a laptop running the model on CPU) this can significantly increase conversion time.

  • Remote mode — increase DOCLING_MCP_SERVICE_TIMEOUT (in seconds) to give the remote service enough time to complete. The default is 300.0 (5 minutes).
  • Local mode / MCP client timeout — the MCP client itself may also enforce a request timeout. For example, in LM Studio set the timeout field (in milliseconds) in your mcp.json:
json
{  "mcpServers": {    "docling": {      "url": "http://localhost:8000/mcp",      "timeout": 300000    }  }}

Markdown export

VariableDefaultDescription
DOCLING_MCP_IMAGE_EXPORT_MODEplaceholderHow images are rendered in Markdown output: placeholder (emits <!-- image -->), embedded (base64 data-URI), referenced (file path / URL)

Cache

VariableDefaultDescription
DOCLING_MCP_CACHE_MAX_DOCUMENTS10Maximum number of documents to keep in the in-memory cache. When the limit is reached the least-recently-used document is evicted.

LlamaIndex RAG (--tools llama-index-rag)

VariableDefaultDescription
DOCLING_MCP_LI_API_BASEhttp://127.0.0.1:1234/v1OpenAI-compatible LLM endpoint
DOCLING_MCP_LI_API_KEYnoneAPI key for the LLM endpoint
DOCLING_MCP_LI_MODEL_IDibm/granite-3.2-8bLLM model identifier
DOCLING_MCP_LI_EMBEDDING_MODELBAAI/bge-base-en-v1.5HuggingFace embedding model

LlamaStack (--tools llama-stack-rag / --tools llama-stack-ie)

VariableDefaultDescription
DOCLING_MCP_LLS_URLhttp://localhost:8321LlamaStack server URL
DOCLING_MCP_LLS_VDB_EMBEDDINGall-MiniLM-L6-v2Embedding model for vector DB
DOCLING_MCP_LLS_EXTRACTION_MODELopenai/gpt-oss-20bModel used for structured extraction

Setting variables in an MCP client config

json
{  "mcpServers": {    "docling": {      "command": "uvx",      "args": [        "--from=docling-mcp",        "docling-mcp-server"      ],      "env": {        "DOCLING_MCP_CONVERSION_MODE": "remote",        "DOCLING_MCP_SERVICE_URL": "https://your-docling-service.example.com",        "DOCLING_MCP_SERVICE_API_KEY": "your-api-key-here"      }    }  }}

Getting started

The easiest way to install Docling MCP and connect it to your client is by launching it via uvx.

Depending on the transfer protocol required, specify the argument --transport, for example

  • stdio used e.g. in Claude for Desktop and LM Studio

    sh
    uvx --from docling-mcp docling-mcp-server --transport stdio
  • sse used e.g. in Llama Stack

    sh
    uvx --from docling-mcp docling-mcp-server --transport sse
  • streamable-http used e.g. in containers setup

    sh
    uvx --from docling-mcp docling-mcp-server --transport streamable-http

More options are available, e.g. the selection of which toolgroup to launch. Use the --help argument to inspect all the CLI options.

For developing the MCP tools further, please refer to the Developing section of CONTRIBUTING.md for instructions.

Integration with MCP clients

One of the easiest ways to experiment with the tools provided by Docling MCP is to leverage an AI desktop client with MCP support. Most of these clients use a common config interface. Adding Docling MCP in your favorite client is usually as simple as adding the following entry in the configuration file.

json
{  "mcpServers": {    "docling": {      "command": "uvx",      "args": [        "--from=docling-mcp",        "docling-mcp-server"      ]    }  }} 

When using Claude for Desktop, simply edit the config file claude_desktop_config.json with the snippet above or the example provided here.

In LM Studio, edit the mcp.json file with the appropriate section or simply click on the button below for a direct install.

[Add MCP Server docling to LM Studio]

Other integrations are described in the integrations page.

Examples

Converting documents

Example of prompt for converting PDF documents:

prompt
Convert the PDF document at <provide file-path> into DoclingDocument and return its document-key.

Generating documents

Example of prompt for generating new documents:

prompt
I want you to write a Docling document. To do this, you will create a document first by invoking `create_new_docling_document`. Next you can add a title (by invoking `add_title_to_docling_document`) and then iteratively add new section-headings and paragraphs. If you want to insert lists (or nested lists), you will first open a list (by invoking `open_list_in_docling_document`), next add the list_items (by invoking `add_list_items_to_list_in_docling_document`). After adding list-items, you must close the list (by invoking `close_list_in_docling_document`). Nested lists can be created in the same way, by opening and closing additional lists.
During the writing process, you can check what has been written already by calling the `export_docling_document_to_markdown` tool, which will return the currently written document. At the end of the writing, you must save the document and return me the filepath of the saved document.
The document should investigate the impact of tokenizers on the quality of LLMs.

Contributing

We welcome external contributions. See CONTRIBUTING.md for details on how to get started.

License

The Docling MCP codebase is under MIT license. For individual model usage, please refer to the model licenses found in the original packages.

LF AI & Data

Docling and Docling MCP is hosted as a project in the LF AI & Data Foundation.

IBM ❤️ Open Source AI: The project was started by the AI for knowledge team at IBM Research Zurich.

来源:README.md,提交 91e7c4f

工具

0
SourceWeft 曾在隔离沙箱中以无凭据方式启动该软件包一次,以列出其工具,但未能完成:该软件包类型无法在沙箱中运行。

版本历史

5
  1. v3.5.0最新Nov 4, 2026
  2. v3.4.0Oct 29, 2026
  3. v3.3.0Oct 6, 2026
  4. v3.2.1Sep 24, 2026
  5. v3.2.0Sep 16, 2026

评分

已满足适用检查项的 50%

该评分衡量此收录条目的完整程度,依据注册中心条目、端点与公开来源计算得出。它不是对服务器本身的评价,也不表示服务器安全或当前可用:可连通性是带日期的检查结果,评分也不包含安装量或用户评价。

  • 近 30 天内可连通25 分

    请保持端点能够响应 MCP 握手;SourceWeft 每 30 天探测一次。

  • 已声明认证方式10 分

    请在注册中心条目中声明凭据请求头,或保持端点能够响应探测,以便判定其认证方式。

  • 已列出工具15 分

    请让服务器在无凭据时响应 tools/list,或发布可在沙箱中无凭据启动的软件包。

  • 描述必需15 分
  • 图标10 分
  • 文档15 分
  • 许可证10 分