Bank statements, payslips, brokerage reports — drop into LangChain, LlamaIndex, or any MCP-aware agent in three lines. Schema-validated, signed webhooks, multi-region. Self-serve from €29/mo.
What you ship
We use Claude with vision + tool-use, so the model emits JSON that's already validated against a strict schema. Then we run a balance-chain check (opening + Σtransactions = closing) to catch sign errors before you ever see them.
EU IBAN, US ABA, UK sort code. Single page or 30 — same endpoint.
Gross, net, deductions, employer details. Schema-validated tax buckets.
Holdings, trades, dividends across multiple custodians.
Line items, VAT, supplier metadata. EU + US tax flavors.
{
"schema_version": "bank_statement.v1",
"model": "claude-sonnet-4-6",
"warnings": [],
"result": {
"institution_name": "NORDIC TRUST BANK",
"account_holder": "Test Customer",
"account_iban": "BE12345678901234",
"statement_period_start": "2026-04-01",
"statement_period_end": "2026-04-30",
"currency": "EUR",
"opening_balance": { "amount": "1000.00", "currency": "EUR" },
"closing_balance": { "amount": "2159.59", "currency": "EUR" },
"transactions": [
{
"posted_date": "2026-04-03",
"description": "Salary payment",
"amount": { "amount": "2500.00", "currency": "EUR" },
"balance_after": { "amount": "3500.00", "currency": "EUR" },
"type": "credit"
},
{
"posted_date": "2026-04-05",
"description": "Rent",
"amount": { "amount": "-1200.00", "currency": "EUR" },
"balance_after": { "amount": "2300.00", "currency": "EUR" },
"type": "debit"
},
{
"posted_date": "2026-04-08",
"description": "Grocery store",
"amount": { "amount": "-87.50", "currency": "EUR" },
"balance_after": { "amount": "2212.50", "currency": "EUR" },
"type": "debit"
}
]
}
}Real response from a synthetic statement · 7 transactions · 0 warnings
Why parsr
We started parsr because every doc-parsing API for a fintech project either failed compliance, started at $500/mo, or both. The opinion baked into the platform: developer-first, EU-first, no-bullshit.
01 / Sovereignty
Compute on Exoscale Zürich (Swiss-based, A1 Group, no US parent). Storage on Cloudflare R2 with jurisdiction='eu'. Postgres + Stripe records on EU instances. DORA, BaFin, and public-sector buyers can tick the box on the first call.
Zero US Cloud Act exposure for EU customers.
02 / Pricing
€19/mo for 1k pages. €99 for 10k. €499 for 100k. Free tier with 100 pages, no credit card to try. Same accuracy as the enterprise vendors charging $500/month entry.
Linear, transparent overage. No quotes.
03 / Accuracy
Most APIs train per-template and fail on the long tail. We force Claude to emit JSON against a strict Pydantic schema, then run a balance-chain validator (opening + Σtransactions = closing). Most layouts work first try; the rest surface as warnings, not silent corruption.
Schema mismatch → 4xx, never wrong data.
04 / Speed
Async with Idempotency-Key by default. Pass `?wait=60` for synchronous mode if you'd rather block than poll. HMAC-signed webhooks on completion with 6-attempt exponential backoff. Region-bound API keys so EU traffic never crosses the Atlantic.
p50 2.1s · p95 4.8s on bank statements.
Who's using it
You don't need a vendor evaluation cycle, a security review, or a minimum commitment. Most teams ship a prototype on the free tier in an afternoon.
Lending & underwriting
AI bookkeeping & accounting
PFM & wealth tools
How it works
Spec lives at docs.tryparsr.dev. Postman collection and SDK examples included.
Send `document_url` or `document_base64`. Add an `Idempotency-Key` header to make retries safe. SSRF-protected: only `https://` URLs to public hosts, no peeking at metadata services.
Worker pulls from NATS JetStream, calls Claude with vision + a JSON schema as a tool, validates the result, computes balance-chain checks, stores in regional R2.
Three ways: poll `/v1/jobs/{id}`; pass `?wait=60` for synchronous mode; or register an HMAC-signed webhook. Whichever fits your flow.
Drop into any stack
Same endpoint shape across every SDK we publish. Idempotency-Key on retries, region-bound bearer token, JSON in and out — the way you already write integration code.
curl https://eu-api.tryparsr.dev/v1/parse?wait=60 \
-H "Authorization: Bearer $PARSR_KEY" \
-H "Idempotency-Key: $(uuidgen)" \
-H "Content-Type: application/json" \
-d '{
"document_url": "https://example.com/statement.pdf",
"doc_type": "bank_statement"
}'“We rebuilt our document onboarding around parsr in two days. Same accuracy as the vendor we were paying $700/month, schema-validated by default.”
“EU residency was non-negotiable for our DORA filing. parsr was the only API that didn't require six rounds of legal back-and-forth.”
“Idempotency-Key + signed webhooks meant zero retry plumbing on our side. Felt like Stripe-tier developer experience.”
Anonymised feedback from design partners during private beta · Public reference logos coming with the launch cohort.
Pricing
One unit = one parsed page. Cached re-uploads are free. Validation-failed parses don't bill. Sandbox keys are unlimited. Upgrade or downgrade at any time.
200 pages/mo + 500 signup-bonus
≈ $32 USD
1,000 pages/mo
≈ $109 USD
5,000 pages/mo
≈ $439 USD
25,000 pages/mo
100K+ pages, committed
Every plan includes EU sovereignty, region-bound keys, HMAC-signed webhooks, idempotency, and unlimited sandbox-mode keys for development.
Security & compliance
Every line below is a real implementation choice — not aspirational. Full security posture at tryparsr.dev/security.
Frequently asked
Everything below is answered the same way in our public docs and in our DPA. No double-speak between sales and engineering.
Bank statements (GA), payslips and brokerage statements (Beta). Invoices and receipts are next on the roadmap. The schema for each type lives in the public docs and is versioned (e.g. bank_statement.v1).
EU keys (sk_eu_…) hit eu-api.tryparsr.dev, which runs on Exoscale Zürich. Documents land in Cloudflare R2 with jurisdiction='eu'. Postgres + Stripe records are EU-hosted. US keys (sk_us_…) hit us-api.tryparsr.dev on Hetzner Ashburn with R2 location='enam'. The two regions never share data.
Median ~2 seconds for a single-page bank statement, p95 ~5 seconds. Multi-page documents scale roughly linearly. You can pass ?wait=60 to block synchronously, or use the default async + webhook pattern.
Send an Idempotency-Key header on POST /v1/parse. We dedupe on (org_id, key) for 24 hours. Cache hits don't pay the LLM cost but DO bill against quota — we treat the unit as a 'parse you initiated', not a 'cache miss we paid for'.
We POST to your URL with a Parsr-Signature header in the format t=<unix>,v1=<hex>. The HMAC is over `<unix>.<raw_body>` using the secret you receive at registration. 5-minute replay window. Reference implementations in Python, Node, Go, and FastAPI in the docs.
Failures with our fault (LLM down, R2 down) don't bill. Customer-fault failures (oversized file, invalid base64) emit a quantity=0 usage event so they show up in your dashboard but don't charge. The job ends in status 'failed' with a structured error_code.
Self-hosted is roadmap, not today. Private regions are available on enterprise — contact us with your volume estimate and required jurisdiction.
99.5% on Starter and Growth, 99.9% on Scale. Enterprise SLA on request. Public status page at status.tryparsr.dev.
Sign up, get a region-bound API key, and parse 100 pages on the free tier. No credit card. No sales call. The first parse is usually live inside 60 seconds.