space ocr Review (2026): Pricing, Features & Honest Verdict

space ocr is a document OCR API that verifies its own output, returning every extracted value with the coordinates it was read from and a flag when the cross-check fails. The verification works and nobody else ships it this cleanly. It also costs 12 times what raw OCR costs. Best for: developers processing receipts and invoices where a wrong number is expensive. Price: From $0.05/page (free plan: yes). Rating: 7.5/10

What is space ocr?

space ocr turns photos of documents into structured, queryable data, and then checks its own work. You send a receipt, an invoice or a form to one of three endpoints and get back typed fields, markdown or plain text. What makes it different is the fourth thing that comes back: every value carries the vertices on the source image where it was read, plus a flag saying whether the value survived a cross-check against what the OCR engine originally saw.

We found it on Product Hunt on 4 August 2026, ranked 12th with 90 upvotes. It ships as three things at once: a REST API, a no-code console where you drag photos into folders that become spreadsheets, and an MCP server so Claude or ChatGPT can file documents and query them back. It is the work of one maker, Yongha Hwang, and it is a Korean and Japanese product first, with the site localised across five languages and checkout billed in JPY.

The problem it targets is specific and real. Every OCR pipeline built on a language model hallucinates occasionally, and the failure is silent. You get a number back, it looks like a number, and nothing tells you the model invented it.

What Are space ocr’s Key Features?

A five-step verification pipeline that catches hallucinations

This is the product. OCR reads the page first and produces coordinates, which is the only step that generates real positional truth. A model then reads the image and that text together to work out what the page means. Every value the model returns gets anchored back to the exact spot OCR read it from, and where the two disagree, that region is cropped and read again. What still fails comes back flagged as text_verified: false and needs_review.

The model never invents coordinates, which is the architectural decision that makes the whole thing work. A hallucinated value has nowhere real to point, so it fails the anchor step by construction. Their published numbers on an internal corpus of 333 hand-labelled cells: 22 cells fixed by verification against 0 broken, 92.1% of returned coordinates have the page’s own OCR spelling the value inside the box, and a flagged cell is 6.4 times likelier than average to be a real problem, with 3 in 4 flags correct.

Note that they call this an internal regression corpus and not a public benchmark, in their own words, on a page explaining how it was measured. That honesty is worth more than the numbers. Most vendors would have called it a benchmark and hoped nobody asked.

Coordinates as evidence, and a UI built to use them

Because every value knows where it came from, the console can do something most document tools cannot: hover a cell and the exact region of the original photo lights up, and it works in reverse from photo to value. Values that failed the cross-check show up in colour rather than sitting quietly wrong among the correct ones, and you correct them with the original image right beside the field.

If you have ever done manual QA on extracted invoice data, you know the actual cost is not the corrections, it is hunting for which page each suspect number came from. Anchoring collapses that. And “review only what is flagged” is a completely different workload from “review everything because any of it could be wrong”.

Three output formats and a place to keep the results

POST /ocr/fields gives you named, typed fields with line items unfolded into rows. POST /ocr/markdown preserves headings, lists and tables, ready for a doc site or an LLM context window. POST /ocr/text gives the raw paragraph-by-paragraph reading. Results stack into folders you create, and GET /view pulls rows back with where, sort, select and pagination running server-side, while GET /space searches folder names, file names, memo bodies and the cell values themselves.

That last part matters more than it sounds. Most OCR APIs hand you JSON and leave storage as your problem. Here there is no database to stand up: create a folder, let rows pile up, query them back with filters. For a small team processing a few thousand receipts, that removes an entire piece of infrastructure.

An MCP server that lets an agent do the filing

Connect one MCP endpoint and an agent gets four tools: create folders and sheets, upload photos as rows, query with filters, and delete after confirming. Hand it a stack of photos and it builds the structure the documents call for, then retrieves only the rows you ask about later. It works with Claude, ChatGPT, Perplexity, VS Code and Cursor.

This is the right shape for agent tooling. The agent does not have to hold the documents in context, it just files and fetches, which is the difference between a workflow that scales past a handful of pages and one that does not.

How Much Does space ocr Cost?

space ocr starts free with 100 credits a month, where one credit is one page processed. The free tier includes 3 sheets and 1GB of storage, resets monthly, and CSV export costs 1 credit. Beyond that, pay-as-you-go is $0.05 per credit with the same per-page price whether you process in the console or through the API, and you are refunded if nothing comes back. Starter is $25 a month for 400 credits (300 plus the 100 free), 15 sheets and 10GB storage with free downloads. Pro is $49 a month for 1,100 credits (1,000 plus 100 free), unlimited sheets and 100GB storage. Annual billing takes 20% off. Everything is billed in JPY at checkout regardless of the dollar figures shown.

Here is the number you have to reckon with. At $0.05 a page, space ocr is $50 per 1,000 pages. Mistral OCR 4 is $4 per 1,000 pages, halved to $2 with the batch discount. Google’s Enterprise Document OCR processor is $1.50 per 1,000. On raw text extraction, space ocr is 12 to 33 times more expensive, and that is not close.

The fair comparison is different, though. Nobody uses raw OCR for structured invoice extraction. Google’s Custom Extractor and Form Parser are $30 per 1,000 pages, and AWS Textract’s forms API is $50 per 1,000 with its purpose-built expense API at $8 to $10. Measured against structured extraction, which is what this actually competes with, $50 per 1,000 is at the top of a normal range rather than outside it. You are paying roughly a Textract-forms price and getting verification and storage that Textract does not include. Whether that is worth it depends entirely on what a wrong number costs you.

Who is space ocr Best For?

Use space ocr if you process documents where a silently wrong value is expensive. Expense reconciliation, invoice ingestion, anything feeding an accounting system or a compliance record. The verification flag is the feature, and it only pays for itself when catching an error is worth more than the per-page premium. If a bad number costs you an hour of somebody’s time, the maths works at almost any volume.

Use it too if you are wiring documents into an agent workflow and do not want to build the storage layer. The MCP server plus server-side query means you can go from photos to a queryable table without provisioning anything, and the 100 free pages a month are enough to prove the concept properly before you spend.

Skip space ocr if you are doing bulk text extraction where occasional errors are tolerable. Digitising an archive, feeding a search index, building a corpus: use Mistral OCR at $2 to $4 per 1,000 pages and save 90% of the cost, because you are paying for a guarantee you do not need. Skip it too if you need volume pricing, since there are no tiers above Pro and no published enterprise rate, so at 50,000 pages a month you are on pay-as-you-go at $2,500 with nothing to negotiate against.

Best space ocr Alternatives

Mistral OCR is the price leader at $4 per 1,000 pages for OCR 4, or $2 with the 50% batch discount, and OCR 3 cheaper still. Quality on clean documents is excellent and markdown output is strong. There is no verification layer and no storage, so you get the reading and nothing else. For high-volume, error-tolerant work it is the obvious choice.

AWS Textract is the enterprise default. $1.50 per 1,000 pages for plain text, $15 for tables, $50 for forms, with the AnalyzeExpense API for receipts and invoices at $8 to $10 per 1,000. It has confidence scores and bounding boxes, which is the closest thing in the mainstream to what space ocr does, though it does not run a second adversarial read. Pick it if you are already on AWS and want a supported service with real volume pricing.

Google Document AI sits similarly: $1.50 per 1,000 for enterprise OCR, $30 per 1,000 for the Custom Extractor and Form Parser, dropping to $20 above a million pages. It is the strongest option for training a custom extractor on your own document types, which space ocr does not offer at all. The setup cost is a GCP project and real configuration work.

Final Verdict: Is space ocr Worth It?

If your documents carry numbers that matter, yes. The verification pipeline is a real engineering idea rather than a marketing one, and anchoring every value to coordinates the model cannot fabricate is the correct way to make LLM extraction auditable. Shipping the API, a usable console and an MCP server together means it fits whether you are writing code or handing the job to an agent. The 100 free pages a month are enough to test it properly on your own worst photos, which is the only test that counts.

The reservations are about the business, not the software. This is one person, the contact address on the site is a personal Gmail account, and checkout bills in JPY, so budget for currency conversion and think about what happens to your pipeline if the maker moves on. There is no volume tier and no enterprise contact, so this does not scale past a certain point on published terms. And the accuracy figures, while unusually honestly labelled as an internal corpus rather than a benchmark, are still self-reported on 333 cells. Run your own documents through the free tier and check the flags against what you already know is in them. If the flagged cells really are the wrong ones, the premium is easy to justify.