# Parse documents without sending them out

> Parse documents without sending them out: 2 companies fully do it, 9 with a limit, 113 carry the tag. Every listing in Document parsing & extraction with the reason and a source.

- Canonical: https://www.anchorterminal.com/niches/document-extraction.private-local-parse/
- Markdown: https://www.anchorterminal.com/niches/document-extraction.private-local-parse/index.md (~1,650 tokens)
- Slim: https://www.anchorterminal.com/niches/document-extraction.private-local-parse/index.min.md (~330 tokens, same facts, less prose, for token-sensitive contexts)
- JSON: https://www.anchorterminal.com/niches/document-extraction.private-local-parse/index.json (this page as data, same URL with Accept: application/json)
- Site index for agents: https://www.anchorterminal.com/llms.txt (the key docs and every listing in one file, at most 16 MB, about 4 million tokens: https://www.anchorterminal.com/llms-full.txt)
- API: https://www.anchorterminal.com/api/v1/index.json
- MCP: https://www.anchorterminal.com/mcp (the whole catalogue as tools, no key)
- DNS: dig +short TXT <slug>.grade.anchorterminal.com (the grade as a DNS TXT record, for a sandbox with no HTTP where names still resolve; also top., best.<category>. and <a>.vs.<b>. under the zone, and the same under grade.letme.dev)
- Licence: CC BY 4.0, credit Anchor Terminal and link this page (https://www.anchorterminal.com/terms/#data)
- Updated: 2026-11-15

Two or three companies' docs show a full fit.

| Companies, full fit | With a limit | Carry the tag | Share of the category's work | First-hand accounts |
| --- | --- | --- | --- | --- |
| 2 | 9 | 113 | 8 | 5 |

## What counts as done

The tool runs on the stated hardware with a documented install, gives the same output as its hosted version where one exists, and the docs state retention, deletion and whether inputs train models.

How agents ask for it: parse these locally, they can't leave the machine; which parser can I self-host; run OCR on my own GPU

## Does the job

The docs show it meets the test above, and nothing we read says otherwise.

- **Docling (Docling Project (LF AI & Data Foundation))**: MIT-licensed library that runs locally on macOS, Linux and Windows, x86_64 and arm64, with a documented CPU-only torch install; there is no hosted version to differ from. acc-18's Windows install failure predates the current install page.
- **Xiaomi-OCR-0 (Xiaomi Corporation)**: Apache 2.0 weights with a documented install; the example code falls back to CPU in float32, and community GGUF builds exist. There is no hosted version to differ from.

## Does the job, with a limit

It does the job, but something in the docs or a user's account will get in an agent's way. The limit is named in each row.

- **Azure Document Intelligence (Microsoft Corporation)**: Read and Layout v4.0 containers run on 8 CPU cores with 10 to 24 GB, but they must send billing data to an Azure endpoint, and hosted inputs are kept 24 hours with a delete call.
- **Datalab (Endless Labs, Inc.)**: Marker, Surya and Chandra weights are free under $2M to $5M revenue thresholds, and acc-14 rated Marker best to self-host, but the supported on-prem container needs an Enterprise contract.
- **Extend API + MCP (CrowdView Inc. dba Extend)**: BYOC deploys into the customer's own cloud VPC and zero retention is available on request, with no training by inference providers, but there is no laptop or single-GPU install with stated hardware.
- **Invofox (Invofox, Inc. (Delaware), with a registered branch in Spain)**: The vendor briefing claims on-premises and VPC deployment and an opt-in zero-retention mode, but no install docs or hardware are published, and corrections feed model training.
- **LlamaParse API + MCP (LlamaIndex, Inc.)**: Self-hosted LlamaCloud runs on Kubernetes through a Helm chart on Enterprise plans only, and still calls OpenAI, Anthropic or Google models, so documents leave the cluster unless enterprise cloud models are set up.
- **Nanonets API + MCP (Nano Net Technologies Inc.)**: The open-source docstrange library runs locally, but acc-41 found local output much worse than the hosted version.
- **Nemotron 3.5 Parse (NVIDIA Corporation)**: Open weights under OpenMDW-1.1 with documented Transformers and vLLM serving, but the model is built for NVIDIA GPUs with no CPU path documented.
- **Reducto API + MCP (Reducto, Inc.)**: Dedicated VPC and on-prem deployments are documented with their own changelog and security model, but they are enterprise contracts, and zero data retention is Growth and Enterprise only.
- **Unstructured API + MCP (Unstructured Technologies, Inc.)**: The open-source library installs with pip or runs from multi-platform Docker images for x86_64 and Apple silicon, but CPU and GPU requirements and any differences from Transform's output are not documented in the README.

## Couldn't check

The docs were closed to us, or silent. These carry no mark either way.

- **Veryfi API + MCP (Veryfi, Inc.)**: No self-hosted option appears in the SDK README, and the retention and training pages were not re-read this pass.

- not this job: Adobe PDF Services / PDF Extract API (Adobe Inc.): Hosted only. Assets are kept 24 hours with a delete endpoint, but nothing runs on the user's machine.
- not this job: Amazon Textract (Amazon Web Services, Inc.): Hosted AWS service only; no self-hosted path.
- not this job: Clairic (Cloudflare, Inc.): Hosted on Workers AI only; open weights are promised with no date. The model page does state zero data retention.
- not this job: Google Cloud Document AI (Google LLC): Hosted Google Cloud service only; no self-hosted processor.
- not this job: LandingAI Agentic Document Extraction (LandingAI, Inc.): Hosted in the US or EU only; the docs index lists no self-hosted or on-premises install. Zero Data Retention is a Team and Enterprise setting.
- not this job: Mindee API (Mindee SAS (Mindee Inc. outside the Eurozone)): Hosted only. Extracted data is kept 12 hours by default (24 at most), the original file is never stored to disk, and processing can be pinned to the EU.
- not this job: Mistral OCR API (Mistral AI (RCS Paris 952 418 325)): Hosted API only; no self-hosted OCR model in the docs read. Zero data retention is available to eligible organisations.
- not this job: OpenDocRouter (LlamaIndex, Inc.): Hosted router only. Without cache nothing is stored, and cached results last 24 hours with a DELETE, but pages go to third-party model providers.

6 listings in the category were added after this reading and aren't marked yet.

## What stops agents finishing

- needs a CUDA GPU the user lacks
- local model weaker than the hosted one with the same name
- parsed output kept by the vendor with no delete

## Around it

Before: file-storage.agent-upload. After: vector-search.local-embedded, document-extraction.rag-ingest.

## Where this comes from

Weight 8 of 100 in Document parsing & extraction: 5 of 46 accounts (Docling, Marker, docstrange, LlamaParse retention); the Docling and Nemotron listings are built for it

As data: https://www.anchorterminal.com/api/v1/use-cases/document-extraction.private-local-parse.json, or the MCP tool get_use_case with id document-extraction.private-local-parse. Draft, read 2026-11-15 from public docs and first-hand accounts, not yet reviewed by a person.

---

Query this catalogue directly over MCP: https://www.anchorterminal.com/mcp (Streamable HTTP, no key). How: https://www.anchorterminal.com/agents/#mcp
