Agnos vs Marker

If your PDFs include scans, complex tables or equations, or you are building a conversion pipeline on a Mac or a GPU server, Marker is the better pick. Its code is free, and its model weights are free for research, personal use and companies under Datalab’s revenue and funding limits that don’t compete with Datalab. If you are on Windows, would rather not install Python, and want folders of text-based PDFs, Office files and web pages cleaned for an AI tool, with a list of what was removed and the rule that removed each piece, Agnos is made for that job.

Gidion makes Agnos and wrote this page, so read it with that in mind. Every statement about Marker comes from Datalab’s own website, documentation, pricing page and GitHub repository, checked on 17 October 2026, and describes Marker 2.0.0, released on 20 July 2026. Each source is listed under Sources. Statements about Agnos describe Agnos 2.3.2 and come from our documentation.

At a glance

Where Marker is the better choice

• Your documents are scans or pictures. Marker OCRs scanned pages and reads image files, and --force_ocr re-reads PDFs whose embedded text is bad. Agnos reads text-based PDFs only and does not read text out of pictures, so a scan gives it nothing to work with.
• Tables, equations and forms matter. Marker formats tables, forms, equations, inline math, links, references and code blocks, and writes equations as LaTeX. Its optional LLM mode, --use_llm, can merge tables across pages and pull values out of forms. Agnos reads PDF tables only in its slower Layout mode.
• The layout is complex, or the pictures matter. Marker detects each page’s layout before it puts the text together, and it extracts and saves the images. Its README notes that very complex layouts, with nested tables and forms, may not work and that forms may not render well, and says --use_llm and --force_ocr mostly solve this. Agnos keeps the text and leaves pictures out.
• Answers must point at the place on the page. Marker’s JSON gives each block its page, its outline on the page and the headings it sits under. Its chunks output does the same in one flat list, which its README says makes chunking for RAG easy. Agnos records the pages, slides, sheet or headings of each piece, not where it sits on the page.
• You work on a Mac or a GPU server. Marker runs on Apple Silicon, on CPUs and on NVIDIA GPUs. Agnos runs on Windows only.
• You are building a pipeline. Marker has a Python API, a command line that converts whole folders with parallel workers, a way to split a big job across several GPUs or machines, and a simple API server for small-scale use. Its processors and renderers can be replaced with your own. Agnos is a desktop app, and its command line is not in the Microsoft Store version yet.
• You’d rather not run it yourself. Datalab sells a hosted API. Its README says the API runs Chandra, a model Datalab rates more accurate than Marker, and keeps no data by default, and its pricing page lists a SOC 2 Type II report. If data must stay in-house, its Enterprise plan runs in your own cloud or fully offline.
• The budget is zero. Marker’s code is free, and its model weights are free for research, personal use, and companies with no more than $5M in revenue last year and $5M in total funding, unless you or your employer offers something that competes with Datalab. Agnos costs $99 once after its 14-day trial.

Datalab publishes its own benchmark scores and speeds for Marker, measured on olmOCR-bench, in the README; those are Datalab’s figures. Agnos publishes no accuracy figures.

Where Agnos is the better choice

• You want no code and no setup. Agnos comes from the Microsoft Store and opens as a window. There is no Python, PyTorch, Docker or model server to install: you start a project and drop in files or a whole folder.
• You want to see what was removed, and why. In Review, each removed piece names the rule that removed it, such as “A running header, footer or page number at the edge of a page”. If cleaning makes more than 500 removals in one file, Review counts them all but lists the first 500. Some pieces can be put back one at a time; for the rest, you switch the rule off and clean again. The Everything export records the removals, with their reasons, in rejected.jsonl. Marker’s README doesn’t describe a list like this.
• You need cleaning, not only converting. Besides page numbers and running headers and footers, Agnos removes cookie banners, “enable JavaScript” notices, placeholder text and passages repeated within a file, each with a named rule in Review. Its Article mode drops a web page’s navigation as the page is read, and an export can keep just one copy of a passage found in several files; the Everything export lists those repeats in duplicates.jsonl. See Cleaning settings and Repeated text.
• Your folders are mixed. One project can hold text-based PDFs, Word, PowerPoint and Excel files, saved web pages, CSV, JSON, Markdown and plain text, cleaned with the same settings and exported together. CSV, JSON and plain text are not among the input types in Marker’s README.
• You want files ready for an AI tool. Agnos exports a RAG corpus with an id and the source on every piece, and its pages, slides, sheet or headings where the file has them; one clean .txt per document for Custom GPTs, NotebookLM, Claude Projects or Copilot; or training data with a validation split that keeps files sharing a passage on the same side. See Export formats.
• Nothing may leave the PC. Marker can run entirely on your own machine too, so the difference is at the edges: its optional LLM mode calls the service you set up (Gemini by default, or a local model through Ollama), and Datalab’s API is hosted, though its README says it keeps no data by default. Agnos sends nothing about your documents anywhere; see Your files and privacy.
• You want one price, paid once. $99 once for one person or $399 once for up to five people at one company, whatever the company’s revenue or funding, with every 2.x update included. No per-page charge and no subscription.

Using them together

If only some of your PDFs are scans, you can use both: Marker reads the scans, and Agnos cleans the result together with the rest of your documents.

1. Convert the scans with Marker. Run marker on a folder of scanned PDFs, or marker_single on one file, with --disable_image_extraction, so it saves no images and its Markdown has no image links. Markdown is its default output. Reading scans needs its model server, described under At a glance.

   marker scans --output_dir markdown --disable_image_extraction

2. Add the Markdown files to an Agnos project. Agnos reads .md files as text, and each piece keeps the headings it sits under. Add them alongside your other documents, such as Word files and saved web pages, in a project with the RAG / Retrieval Corpus purpose.

3. Review and export. Review lists what Agnos removed from each file, such as web clutter and passages repeated within it. Then export a RAG corpus, cleaned .txt files or training data. By default, the RAG corpus and training data keep one copy of text found in several files, and the Everything export lists those repeats in duplicates.jsonl.

Three things to know. Marker writes each document to its own folder, with a _meta.json file beside the .md, so add only the .md files, not the whole output folder. In File Explorer, a search for *.md in the output folder finds them all, ready to drag onto the drop box. Markdown has no pages, so the pieces Agnos makes from it carry headings but no page numbers. And Marker removes page headers and footers by default, so expect few of those in what Agnos receives from it.

In short

If Marker already does what you need, keep using it. If you want folders of documents cleaned on Windows without code, with a list of what was removed, you can try Agnos free for 14 days, with no card. It comes from the Microsoft Store and runs on Windows 10 and 11.

Sources

Marker and Datalab, checked 17 October 2026:

• Marker’s README on GitHub: features, the managed platform, installation and the model server it needs, modes, options, output formats, hybrid LLM mode, limitations, commercial usage and deployment.
• Marker’s source code: one output folder per document with a _meta.json file (output.py, config/parser.py), the page and outline of each block in JSON and chunks (renderers/json.py, renderers/chunk.py), and no image links with --disable_image_extraction (renderers/html.py).
• Marker’s LICENSE: the Apache License 2.0.
• Marker’s MODEL_LICENSE: the modified AI Pubs Open Rail-M licence for the model weights, with its revenue, funding and competitor limits (Attachment A) and its attribution terms (section 7).
• Marker 2.0.0 release notes, released 20 July 2026: the modes, and full CPU support with no model server.
• marker-pdf on PyPI: version 2.0.0 and the Python requirement.
• Datalab’s home page: languages covered and where its models can run.
• Datalab’s pricing: plans, per-page rates, the usage Team includes, the SOC 2 Type II report, Enterprise in your own cloud or fully offline, and who needs to pay for the models.
• Datalab’s documentation: the hosted API and its open-source tools.
• Datalab’s public playground.

Agnos:

• Supported files, Cleaning settings, Export, Export formats, Review, Import documents, Your files and privacy, What Agnos doesn’t do and Command line.
• Pricing, Trial and licences, Download and the licence terms.
• The guide PDF to JSONL for RAG and fine-tuning, for the fields Agnos writes and how it splits training data.

Marker is a trademark of its owner. This page is ours, not theirs; tell us at hello@agnos.cloud if anything here is out of date.