Agnos vs Docling: when to use which

If you write code, have scanned pages or complex tables, work on a Mac or Linux, or are building a pipeline or a service, Docling is the better pick, and it is free. If you want a folder of text-based PDFs, Office files and web pages cleaned on Windows without code, with a list of what it removed and the rule for each, Agnos is made for that job.

Gidion makes Agnos and wrote this page, so read it with that in mind. Every statement about Docling comes from Docling’s own website, documentation and GitHub repositories, or from IBM’s page for its paid Docling service, listed under Sources.

At a glance

Compared	Agnos	Docling
What it is	A Windows app, made by Gidion, that cleans documents with named rules before they go into AI tools, and lists what it removed.	An open-source document converter, started by IBM Research Zurich and now hosted by the LF AI & Data Foundation. It reads each document into one structured format, the DoclingDocument, and exports it from there.
Who it’s for	People preparing documents for Custom GPTs, NotebookLM, Claude Projects, Copilot, a RAG index or training data, without writing code.	Teams building RAG and search, agents and automation, or data extraction from Python, Java or TypeScript code, or running conversion as a service.
Where it runs	Your PC. Documents are never uploaded; Agnos goes online only to activate, check or deactivate a licence.	Your computer by default, fully offline once its models have been downloaded on the first run. Or your own server, with Docling Serve in a container. IBM also sells Docling for IBM watsonx: a managed cloud service on AWS, where documents are processed by the provider, not on your own hardware, and, docling.ai says, a private or on-premises deployment.
Needs code	No. It installs from the Microsoft Store, with no Python to set up. Its command line is coming to the Store version in an update.	Python 3.10 or newer, and `pip`, to install it. Then a Python API, a command line, or Docling Serve, an HTTP server with an optional browser interface. IBM’s paid service has its own web interface.
File types	Text-based PDF, Word (`.docx`), PowerPoint (`.pptx`), Excel (`.xlsx`), web pages, CSV, JSON, plain text, Markdown and source code. Old Office formats, OpenDocument, images, audio, video and archives are refused.	PDF, Word, PowerPoint and Excel, their older formats and RTF through LibreOffice, OpenDocument, Apple Pages, Numbers and Keynote, HTML, Markdown, AsciiDoc, LaTeX, EPUB, CSV, email, images, audio and video, and several XML formats. Some need an add-on.
Scanned PDFs (OCR)	No OCR. A page with no text layer is skipped and named in the file’s warnings; a PDF with no text at all is refused.	Yes. OCR is on by default in its command line, with a choice of engines, among them EasyOCR, Tesseract, RapidOCR and OcrMac. Vision language models such as GraniteDocling can read pages too.
Tables	Read from Word, PowerPoint and web pages, and the RAG / Retrieval Corpus purpose keeps them as written. In PDFs, only with the slower Layout reading mode. Excel sheets become tab-separated rows, without formulas or formatting.	A table structure model recovers rows, columns, multi-level headers and merged cells, and keeps a table’s caption with it. Its docs say JSON and HTML keep merged cells, and Markdown flattens them.
What it outputs	A RAG corpus in JSONL (each piece with an id, its source file and its pages, slides, sheet or headings), cleaned `.txt` files, training data with a validation split, or all of these with a record of what was removed.	Markdown, HTML, lossless JSON, YAML, plain text, DocTags, DocLang, WebVTT and LaTeX, and chunks for RAG in JSONL. Where a document has pages, each element keeps its page and its position on it.
Shows what it removed, and why	Yes. Review lists what it removed and the rule for each, up to 500 removals in a file, and the Everything export adds `rejected.jsonl`, with every removed piece and its reason.	It labels what it finds, such as page headers, footers, captions and tables, and keeps headers and footers in a separate layer from the body. By default its exports include only the body, so headers and footers are left out; the lossless JSON keeps them. It can also grade each conversion, from poor to excellent.
Price	A 14-day trial of everything in Pro, no card. Then $99 once for one person on up to 2 devices, or $399 once for up to five people at one company, on up to 10 devices. Tax is added where it applies. No subscription.	Free. IBM’s managed service is priced in resource units of 1,000 pages: from US$4 a unit, pay as you go, or from US$3,999.60 a year for 1,000 units, with support, on IBM’s price list.
Operating systems	64-bit Windows 10 (version 2004 or later) and Windows 11.	macOS, Linux and Windows, on x86_64 and arm64.
Licence	Commercial, under an end user licence agreement.	Open source, under the MIT License, copyright The Docling Contributors. The AI models it uses have licences of their own.

Where Docling is the better choice

Your documents are scans or photos. Agnos has no OCR, so a scanned page gives it nothing to read. Docling runs OCR by default in its command line, lets you choose the engine and its languages, and can read pages with vision language models such as GraniteDocling instead.
Layout and tables matter. Docling’s models find the reading order of a page with two columns, recover table structure with merged cells and multi-level headers, turn formulas into LaTeX, and classify pictures. Its README lists chart understanding, which turns bar, pie and line charts into tables or code, with a description. The formula, picture and chart models are options you switch on, such as --enrich-formula; the table model is on by default. It can also pull a defined set of fields out of a PDF or image against a schema you give it. Agnos reads PDF tables only in its slower Layout mode.
Answers must point at the page. In a PDF, each Docling element keeps its page number and its box on the page, so an answer can be traced to the exact place it came from. Agnos records the pages, slides, sheet or headings of each piece, not its position on the page.
Other kinds of file. Docling reads old Office formats and RTF (with LibreOffice installed), OpenDocument, Apple iWork files, EPUB books, email, LaTeX, AsciiDoc, images, and audio and video with speech recognition. Agnos refuses old Office formats, OpenDocument, images, audio and video; see what Agnos refuses.
You don’t use Windows. Docling works on macOS, Linux and Windows, on x86_64 and arm64. Agnos runs only on 64-bit Windows 10 and 11.
You are building a pipeline or an agent. Docling is a Python library with a command line that converts a whole folder, an HTTP server, Java and TypeScript clients, and an MCP server for AI agents. It plugs into LangChain, LlamaIndex, Haystack, Crew AI and others, and its own chunkers split documents by structure and by token count. Agnos is a desktop app, and its command line is not in the Microsoft Store version yet.
You have a very large volume. Docling Serve can queue work through Redis, Docling Jobkit runs distributed jobs on Kubeflow or Ray, and the documentation covers running the models on a GPU. Agnos cleans one file at a time on one PC.
You want it free and open. Docling is free under the MIT License, so you can read its code, change it, and build it into your own software. Check the licences of the models you use, which its README says are separate.

Where Agnos is the better choice

You don’t write code. Agnos installs from the Microsoft Store and brings everything it needs: no Python, no pip, no models to download. You drop in a folder, check the result and press export. Docling starts with Python 3.10 or newer; its command line, and Docling Serve’s browser interface, need Python or a container set up first. IBM’s paid managed service has a web interface with nothing to install, but there your documents are processed by IBM, not on your PC.
Your documents never leave your PC, with nothing to configure. Agnos reads, cleans and exports on your computer, has no account and no online options to switch on, and goes online only to activate, check or deactivate your licence, sending no document text. Docling also runs locally by default, and offline once its models are downloaded; the difference is the setup, not the privacy.
You need the clutter taken out. Agnos’s rules remove navigation menus, cookie banners, “enable JavaScript” notices, placeholder text, page numbers, running headers and footers at the edges of PDF pages and slides, and passages repeated within a file or across files. Docling already leaves page headers and footers out of its exports by default; Agnos’s rules also cover the web and repeated clutter.
You want to see what was removed. Review lists what Agnos removed, with the name of the rule that removed each piece, and Keep this anyway puts a removed piece back. Lines such as page numbers and cookie notices come back only by switching their setting off and cleaning again. The Everything export adds rejected.jsonl, with every removed piece and its reason. If one file has more than 500 removals, Review counts them all but lists the first 500.
You have a folder of mixed files for an AI tool. One project takes text-based PDFs, Word, PowerPoint and Excel files, saved web pages, CSV and JSON together, and cleans each one as it arrives. Agnos then exports a RAG corpus in JSONL, cleaned .txt files for Custom GPTs, NotebookLM, Claude Projects or Copilot, or training data split by file into training and validation sets.
You’d rather pay once than build and look after a pipeline. Docling is free, and IBM’s managed version is priced per page. Agnos is $99 once for one person, or $399 once for up to five people at one company, after a 14-day trial with no card. There is no subscription, and every 2.x update is included.

Using them together

Docling writes Markdown files, and Agnos reads .md files and whole folders, so Docling can go first for the files Agnos cannot read, such as scans. Agnos then cleans the Markdown, lists what it removed and builds the export.

Convert with Docling. With Docling installed, convert a folder to Markdown in a terminal:
```
docling convert .\scans --to md --image-export-mode placeholder --output .\markdown
```
OCR is on by default. Without --image-export-mode placeholder, Docling writes each picture into the Markdown as base64 text; with it, only the picture’s position is marked.
Add the folder to Agnos. In a project with the RAG / Retrieval Corpus purpose, which keeps headings, tables and lists as written, press Choose a folder on the Import step and pick the Markdown folder.
Review and export. Open Review, choose a file and press Removed, then build the export you need. In a RAG corpus, each piece carries the Markdown headings it sits under.

Two things to know:

For a text-based PDF, you can give Agnos the PDF itself, and its pieces keep their page numbers. Go through Docling first for scans, or when a PDF’s tables need Docling’s table model. Markdown has no pages, so Agnos’s rule for headers and footers at the edges of a page has nothing to look at; Docling has already left them out by default. Agnos still removes repeated passages, page markers on a line of their own and the other clutter its rules cover.
Both steps keep your documents on your computer. Docling needs the internet on its first run, to download its models; after that it runs offline. If you use IBM’s managed Docling service instead of the local one, your documents leave your computer at that step.

Both tools are good at their own job. Docling turns many kinds of document, including scans, into structured data for code, pipelines and AI agents; Agnos cleans a folder of documents on Windows and shows what it took out. If yours is the second job, you can try Agnos free for 14 days, with no card. What it does not do is listed in What Agnos doesn’t do.

Sources

Checked 17 October 2026. Docling’s website has no pricing page: the open-source project is free, and the price of IBM’s managed service is on IBM’s product page.

Docling

docling.ai: what Docling converts, the MIT licence, running fully offline, Python 3.10 or newer, models downloaded once and cached, reading order, table structure with spans and captions, formulas, pictures, page and position for each element, schema-based extraction, and IBM’s commercial offering, managed or private and on-premises.
Use cases: RAG and enterprise search, agents and automation with schema-based extraction, research and technical documents, and enterprise archives.
Supported formats on docling.ai: what Docling reads and writes, and the parts it labels, such as headers and footers.
Deployments: the local library, Docling Serve, the managed service where the provider processes documents, private or on-premises deployment, and model downloads on the first run.
Ecosystem: integrations, and the Java and TypeScript clients.
docling-project/docling on GitHub: the README’s features, chart understanding, the systems it works on, the start at IBM Research Zurich, hosting by the LF AI & Data Foundation, and the note on model licences.
LICENSE: the MIT License, copyright The Docling Contributors.
Changelog: version 2.140.0, the latest, from 16 October 2026.
Installation: macOS, Linux and Windows on x86_64 and arm64, the add-ons, and the OCR engines.
Supported formats in the docs: every input and output format, and which need LibreOffice or an add-on.
CLI reference: converting folders, the output formats, OCR and the table model on by default, the formula, picture and chart models off by default, and the image export modes.
Docling Document: the body, and the separate furniture layer for headers and footers.
Docling Document API reference and docling-core’s content layers: exports include only the body layer unless told otherwise.
Serialization: which formats keep merged table cells.
Chunking: the chunkers that split by structure and tokens.
Confidence scores: the grades from poor to excellent.
FAQ: running completely offline, in air-gapped environments.
API server, Jobkit, GPU support and MCP server: running as a service with a Redis queue, distributed jobs on Kubeflow or Ray, GPUs, and AI agents.
docling-project/docling-serve on GitHub: the server, its container images and its browser interface.
Managed service in the Docling docs, and IBM’s Docling for IBM watsonx page: the hosted service on AWS, its web interface, its two price plans, from US$4 per 1,000 pages or US$3,999.60 a year, and the Docling trademark.

Agnos

Docling is a trademark of LF Projects, LLC. This page is ours, not theirs; tell us at hello@agnos.cloud if anything here is out of date.