Extract your data.Every value cited.

Turn invoices, bank statements, contracts, receipts and forms into structured data. Every value points to the exact words it came from and carries a calibrated confidence, so sure values flow straight through and only doubtful ones reach a person.

Free to start: 1,000 welcome credits (Flash tier, about 250 pages, valid 30 days). Then prepaid packs from €10, no subscription. See pricing

Fields extracted from the invoice

invoice-0142.pdf

  1. Ingest: done
  2. Parse: done
  3. Extract: done
  4. Validate: done
  5. Review: waiting for review

Review: 1 field waiting

Fig. 1 — Pick a field: its value points to the exact words it came from. Synthetic invoice and data.
  • Source-available core (FSL‑1.1‑ALv2)
  • Self-host with Compose or Helm
  • Managed service in the EU
  • Your own model keys (BYOK)

Every value, cited word by word

A deterministic verifier checks every value against the document text: the literal, its normalized form or its number. The viewer highlights the exact words behind it, and an ambiguous value is never guessed: it is flagged.

Table 1 — Verification states of a citation
StateWhen
VerifiedThe value, raw, normalized or numeric, is in the cited region; or it was found elsewhere and the citation corrected.
PartialOnly an approximate text match, with a similarity of 0.85 or more.
Block onlyA yes or no, or an option, with no literal in the text, backed by a phrase of the cited region.
ContradictedThe cited region holds another value of the same type: another amount, another date, the opposite option.
UnverifiedAnything else: the document does not back the value.

Only doubtful values reach a person

Each field's confidence is calibrated on your own evaluations into a probability per segment. A certified threshold decides what is approved automatically, a random audit rechecks a share of those approvals, and everything else goes to an inbox with queues, assignment, SLAs and double review.

Calibrated confidence of each field of the invoice

  • supplier0.99, Auto-approved
  • invoice_number0.98, Auto-approved
  • date0.97, Auto-approved
  • total0.98, Auto-approved
  • iban0.62, To review
  • To review
  • Auto-approved
Fig. 2 — A certified threshold separates what is approved automatically from what a person reviews. Synthetic values.

Review inbox

DocumentQueueReasonSLA
invoice-0142.pdfSuppliersLow confidence3 h 20 min
statement-october.pdfBanksBlocking rule faileddue soon
lease-12.pdfContractsRandom audit1 d 4 h
taxi-receipt-0907.jpgExpensesReviewers disagreeoverdue
Fig. 3 — Review inbox: every task says why it exists. Synthetic data.

Visual workflows, and an assistant that builds them

In Studio you chain the steps (parse, classify, split, extract, validate with rules, review, export) and test them on sample documents. The assistant builds or edits the workflow from a conversation, but never changes anything on its own: it proposes a change with its diff, and you approve it. Approving creates a draft; publishing is still your call.

invoices-supplierv14Draft
  1. Parseparsedefault
  2. Classifyclassify3 categories
  3. Splitsplit2 rules
  4. Extractextractinvoice · standard
  5. modifiedValidatevalidate5 rules
  6. Reviewreviewon_low_confidence
  7. Exportexportwebhook erp

Check that the line items add up to the subtotal.

Proposed change: 1 rule added to validate+ arithmetic: sum line_items[*].amount = subtotal

Fig. 4 — Studio: the change the assistant proposes is marked before you approve it. Synthetic data.

For developers and agents

The same cited extraction from your code, your terminal or your agent, behind an API contract built for integration.

REST API
RFC 9457 errors, Idempotency-Key, ETag and signed webhooks.
Python SDK
Pydantic schemas and typed results with their evidence.
TypeScript SDK
zod schemas and the same workflow builder.
CLI
One command, no workflow and no schema file to create.
MCP server
For Claude Code, Codex and Cursor, with approvals and OAuth 2.1.
Code example
$ extractit invoice.pdf --field total:number
{
  "data": { "total": 1234.56 },
  "fields": [
    {
      "path": "total",
      "value": 1234.56,
      "raw_text": "€1,234.56",
      "confidence": { "score": 0.98, "band": "high", … },
      "evidence": [
        {
          "page": 1,
          "quote": "Total due €1,234.56",
          "verification": "verified",
          "words": [ … ]
        }
      ],
      "review_state": "auto_approved"
    }
  ]
}
from extractit_sdk import DX, Tier
from pydantic import BaseModel

class Invoice(BaseModel):
    invoice_number: str
    total: float

r = DX().extract("invoice.pdf", schema=Invoice, tier=Tier.STANDARD)
print(r.data.total)                  # 1234.56
print(r.confidence("total").score)   # 0.98
print(r.evidence("total")[0].quote)  # Total due €1,234.56
import { DX } from "@extractit/sdk";
import { z } from "zod";

const Invoice = z
  .object({ invoice_number: z.string(), total: z.number() })
  .meta({ title: "Invoice" });
const dx = new DX({ apiKey: process.env.DX_API_KEY });
const run = await dx.extract(file, { schema: Invoice });

console.log(run.data.total);                         // 1234.56
console.log(run.fields.total?.evidence[0]?.regions);
$ claude mcp add --transport http extractit \
    https://app.extractit.dev/mcp \
    --header "Authorization: Bearer $DX_API_KEY"

# Codex and Cursor use the same URL: per-client guide in the docs
Fig. 5 — One extraction from the terminal: the value, its quote and its state. Output abridged.

Quality you measure, not assume

Every workflow shows its precision on reviewed fields, its mean confidence and its STP, the share of runs that finish without review. Datasets have sealed splits, evaluations are immutable, and the optimizer tests better field descriptions in measured rounds; only a person promotes the winner.

Workflows

NameAlias → versionDocs 30 dPrecisionConfidenceSTP
invoices-supplierprod v142,34098.1%93%86%
bank-statementsprod v61,11596.4%90%79%
rental-contractsv3 · Draft38No dataNo data—
Fig. 6 — Studio workflow list. Synthetic data.

Your documents, where you decide

The core is source-available (Fair Source, FSL‑1.1‑ALv2; each version becomes Apache‑2.0 after two years): install it with Compose or Helm on your infrastructure, and it can use local models. Model policies (EU only, zero retention, air-gapped, your keys only) decide which provider may see a document and, when none qualifies, the call fails closed.

Table 2 — Editions
FeatureCommunityFSL-1.1-ALv2Commercial editionFSL-1.1-ALv2
Studio, API, review and evaluationsIncludedIncluded
Deployment with Compose and HelmIncludedIncluded
Model policies and your own keys (BYOK)IncludedIncluded
SSO (OIDC and SAML) and SCIM provisioningNot includedIncluded
Air-gapped profileNot includedIncluded

FSL‑1.1‑ALv2 is source-available: each version becomes Apache‑2.0 two years after its release. The SDKs and schemas are Apache‑2.0.

Managed service. If you would rather not run it, the managed service is hosted in an EU region and paid with credits.

Pay for what you process

Prepaid credits, with no subscription or monthly fee. Each page uses the credits of the tier of the step that extracts it, and you choose the tier by how hard the document is.

Table 3 — Credit packs
PackCreditsPricePer 1,000 creditsDiscount
welcomeFlash only · 30 days1,000Free——
pack_10k10,000€10€1.00—
pack_50k50,000€45€0.90−10%
pack_200k200,000€160€0.80−20%

Every new organization gets 1,000 free welcome credits: Flash tier only (about 250 pages), valid for 30 days, one per verified email. There is no monthly free plan; the source-available core (FSL‑1.1‑ALv2) is free and you can self-host it.

Start with one document

ExtractIt is getting ready to launch. Ask for early access and your organization will start with 1,000 free welcome credits. Meanwhile, the quickstart shows the path from an API key to your first cited value and an agent connected over MCP.

  1. Create an API keyIn the app, under Settings → API keys.
  2. Extract an invoiceextractit invoice.pdf -f total:number --remote
  3. Connect your agentPoint any MCP client at app.extractit.dev/mcp.