Every API here is available over the APIs.io API and to AI agents over MCP.
openapi: 3.2.0
info:
title: Document Extraction MCP API
version: 1.0.0
description: 'Email sign-up at /app/api/auth/sign-up/email requires an x-turnstile-token
header in production. Failed or unavailable bot verification returns 403
with code BOT_CHECK_REQUIRED. Social sign-up is unaffected.
Documents in. Verified data out.
The public REST API (v1) turns documents (invoices, receipts, purchase
orders, bank statements, delivery notes, or custom schemas) into
structured, verified data. Every document reaches an explicit terminal
state. Every extracted table row is accounted for. Money fields are
cross-checked (crossfooting). Bank statements are balance-checked.
Every field links back to its source in the document.
Unsupported methods on registered API paths return an empty
405 Method Not Allowed response with an Allow header listing supported
methods. HEAD is supported wherever GET is registered; OPTIONS keeps its
existing preflight behavior. Unknown paths remain 404. Maintenance and
canonical-host redirects still take precedence.'
contact:
email: support@arkrel.com
servers:
- url: https://arkrel.com
description: Production, one Worker, one domain, same origin as the site itself (no separate api.<domain> host)
security:
- ApiKeyAuth: []
tags:
- name: MCP
paths:
/mcp:
post:
tags:
- MCP
operationId: mcpRpc
summary: Remote MCP (Model Context Protocol) server
description: 'Single JSON-RPC 2.0 endpoint for the remote MCP server (stateless
Streamable HTTP, one request in, one response out, no
Mcp-Session-Id, no SSE). Same `Authorization: Bearer `
auth as every other `/v1/*` route. `Content-Type` and `Accept`
must both be `application/json`.
This entry documents the transport shape only, not every JSON-RPC
method. JSON-RPC''s method dispatch doesn''t map onto OpenAPI''s
per-path-per-verb model. The full method/tool reference (
`initialize`, `tools/list`, `tools/call` with `extract_document`,
`get_document_result`, `list_documents`, `get_usage`, their input
schemas, and worked examples) is documented at `/docs/mcp-server`
and implemented in `api/src/mcp/{protocol,handler,tools}.ts`.
`GET`/`DELETE`/`PUT`/`PATCH` on this path return `405` (this
transport revision has no server-initiated stream to open or
close); `OPTIONS` is the CORS preflight.'
requestBody:
required: true
content:
application/json:
schema:
$ref: '#/components/schemas/JsonRpcRequest'
example:
jsonrpc: '2.0'
id: 1
method: tools/call
params:
name: get_usage
arguments: {}
responses:
'503':
$ref: '#/components/responses/MaintenanceMcp'
'200':
description: 'A JSON-RPC response (success result or a JSON-RPC-level error,
an MCP tool error is also carried as a 200 with
`result.isError` true, per the MCP spec, not an HTTP error
status).
'
content:
application/json:
schema:
$ref: '#/components/schemas/JsonRpcResponse'
example:
jsonrpc: '2.0'
id: 1
result:
period_start: '2026-09-01T00:00:00Z'
period_end: '2026-10-01T00:00:00Z'
pages_included: 50
pages_used: 32
pages_overage: 0
'202':
description: Accepted, the body was a JSON-RPC notification (no `id`), which never gets a JSON-RPC response by design.
'400':
description: Malformed JSON, an invalid JSON-RPC envelope, or an unsupported MCP-Protocol-Version.
content:
application/json:
schema:
$ref: '#/components/schemas/JsonRpcResponse'
'401':
description: Missing or invalid API key. Carries `WWW-Authenticate` (unlike the plain `/v1/*` 401) plus a JSON-RPC error body.
content:
application/json:
schema:
$ref: '#/components/schemas/JsonRpcResponse'
'403':
description: Origin not allowed, or the key creator requires TOTP setup. Setup rejection uses JSON-RPC error -32005 with error.data.code TWO_FACTOR_SETUP_REQUIRED and id null, before request dispatch. Keys without a creating user are exempt.
content:
application/json:
schema:
$ref: '#/components/schemas/JsonRpcResponse'
example:
jsonrpc: '2.0'
id: null
error:
code: -32005
message: Set up an authenticator app to continue.
data:
code: TWO_FACTOR_SETUP_REQUIRED
'405':
description: Method not allowed (GET/DELETE/PUT/PATCH). Carries an `Allow` header.
'406':
description: Accept header does not allow application/json.
'413':
description: Request body exceeds the 34 MiB limit.
'415':
description: Content-Type is not application/json.
'429':
description: Rate limited. Carries `Retry-After` and the standard `X-RateLimit-*` headers. Organization budgets are shared with app and /v1 requests. Bookkeeping mutations allow 30 requests per minute, document extraction 60, and general reads 300.
content:
application/json:
schema:
$ref: '#/components/schemas/JsonRpcResponse'
options:
tags:
- MCP
operationId: mcpCorsPreflight
summary: CORS preflight for browser-hosted MCP clients
description: '204 with no CORS headers when there''s no Origin (not a browser
preflight, e.g. a server-side MCP client or curl). With an
allowed Origin (same-origin, or localhost/127.0.0.1 at any port
for local dev tooling like the MCP Inspector), 204 with
`Access-Control-Allow-Origin` echoed back. A disallowed
third-party Origin gets 403, no CORS headers.'
responses:
'204':
description: Preflight allowed (or not a real preflight, no Origin header at all).
'403':
description: Origin present but not allowed.
components:
responses:
MaintenanceMcp:
description: A mutating tool is temporarily unavailable. Read tools remain available.
headers:
Retry-After:
schema:
type: integer
const: 60
content:
application/json:
schema:
$ref: '#/components/schemas/JsonRpcResponse'
example:
jsonrpc: '2.0'
id: 1
error:
code: -32004
message: Arkrel is undergoing maintenance. Please try again in a minute.
schemas:
JsonRpcResponse:
type: object
description: JSON-RPC 2.0 response, exactly one of `result`/`error` is present, matching whichever `id` was in the request.
required:
- jsonrpc
- id
properties:
jsonrpc:
type: string
enum:
- '2.0'
id:
type:
- string
- number
- 'null'
result:
type: object
additionalProperties: true
error:
type: object
required:
- code
- message
properties:
code:
type: integer
description: A JSON-RPC error code (the standard -32xxx range, or an MCP-specific code, see api/src/mcp/protocol.ts).
message:
type: string
data:
additionalProperties: true
example:
jsonrpc: '2.0'
id: 1
result:
tools:
- name: extract_document
description: Upload a document and get back verified fields, tables and checks.
JsonRpcRequest:
type: object
description: JSON-RPC 2.0 request or notification (a notification omits `id`). See `/docs/mcp-server` for the full method/tool reference.
required:
- jsonrpc
- method
properties:
jsonrpc:
type: string
enum:
- '2.0'
id:
description: Omitted for a notification. Otherwise a string or number the response echoes back.
type:
- string
- number
- 'null'
method:
type: string
description: e.g. "initialize", "tools/list", "tools/call".
params:
type: object
additionalProperties: true
example:
jsonrpc: '2.0'
id: 1
method: tools/list
securitySchemes:
BusinessSession:
type: apiKey
in: cookie
name: better-auth.session_token
description: Better Auth session cookie. Mutations also require a same-origin Origin header.
EngineToken:
type: apiKey
in: header
name: X-Engine-Token
description: Shared operator token for internal maintenance.
SessionCookie:
type: apiKey
in: cookie
name: __Host-arkrel_session
description: Signed-in app session. Mutations also require a same-origin request. Sensitive actions by accounts with a password without completed protection receive 403 TWO_FACTOR_SETUP_REQUIRED as described in x-app-account-protection. The free first look and account deletion remain available.
ApiKeyAuth:
type: http
scheme: bearer
description: 'API key shown once at creation, e.g. `sk_live_...` / `sk_test_...`.
Stored server-side as a salted hash; only the prefix is retrievable
afterward.
API keys and MCP credentials are checked against their creating user on every request. If that user has a password without enabled TOTP, requests return HTTP 403 TWO_FACTOR_SETUP_REQUIRED. Organization/service keys with no creating user keep their existing behavior.
'
x-webhook-delivery-payload:
$ref: '#/components/schemas/WebhookDeliveryPayload'
x-app-email-signup:
path: /app/api/auth/sign-up/email
method: POST
header: x-turnstile-token
status: 403
code: BOT_CHECK_REQUIRED
description: A fresh single-use Turnstile token with action signup and hostname matching APP_BASE_URL is required in production. Missing, invalid, expired, or unavailable verification fails closed. Reset the widget after each submission. Google sign-up is unaffected. Local development without a configured secret does not require a challenge. This route is served by the mounted Better Auth handler.
x-app-account-protection:
status: 403
code: TWO_FACTOR_SETUP_REQUIRED
response:
$ref: '#/components/responses/TwoFactorSetupRequired'
description: 'Password accounts require verified TOTP before checkout or billing changes, reports and exports (PDF, CSV, accountant CSV, QuickBooks, Xero, OFX, JSON downloads, statement exports and data ZIP), share creation and use, API keys and MCP, bank or ledger connections, team changes, email/password changes, inbound sender allowlist changes, callback secrets, and uploads with webhook_url. Uploads that would consume paid page credits or record billable overage also return TWO_FACTOR_SETUP_REQUIRED before storage. Included monthly and purchased Catch-Up allowances remain available. Admission reserves pending pages, and settlement rechecks live protection against actual processed pages. Chart-of-accounts sync and all chart writes require protection; stored chart and categorization-settings reads do not contact a ledger provider. Session uploads with a callback return TWO_FACTOR_SETUP_REQUIRED before storing or scheduling the document when the password account needs setup. The same refusal applies when the business has an active saved outbound webhook, even if webhook_url is omitted. Queued webhook dispatch checks live creator and document-uploader protection before sending, including retries. Existing role, entitlement and step-up requirements remain. Linked passkeys or providers do not exempt password accounts. Passwordless accounts keep existing rules. Email verification remains required. Sign-up and sign-in do not redirect to setup. The free first look stays open: one statement upload or email-in, processing, merchant review, rules, My year and locked report previews, billing/pricing views, profile and business name, sign-out, and account deletion with its existing step-up. Sensitive actions prompt inline and resume after setup. Direct /protect-account links remain supported. Other app routes default to protection. Exemptions use explicit methods and complete route patterns; business settings changes other than name-only renames, session revocation, provider linking and token access require protection. Implicit same-email provider linking is disabled: sign in with the password first, protect the account and pass step-up, then link explicitly. Explicit linking and passkey registration require fresh step-up and recheck at credential creation or verification. Already-linked provider sign-in and first-time provider signup remain available. Authentication, recovery and TOTP enrollment ceremonies remain available. Credential mutation routes require protection, including passkey registration options and verification. OAuth and passkey removal also require fresh step-up (401 STEP_UP_REQUIRED). Password credential unlinking is refused with 403 PASSWORD_REMOVAL_NOT_ALLOWED even after enrollment and step-up, preserving the password account protection requirement. The server reads live enrollment. API keys and MCP credentials retain their creator check; creatorless service keys retain existing behavior. Share links created by an unprotected password account remain unavailable.'
open_workflows:
- Free first look and locked report preview
- Billing and pricing views
- Profile name and sign-out
- Account deletion with existing step-up
- Authentication and protection enrollment
session_status:
path: /app/api/auth/me
rate_limits:
window_seconds: 60
per_session: 180
per_ip: 600
status: 429
retry_header: Retry-After
description: Independent read budgets. The IP budget precedes session resolution; the session budget uses the resolved session ID before membership and protection reads. These reads do not consume credential budgets.
field: two_factor_setup_required
schema:
type: boolean
description: Present on authenticated responses. True means sensitive actions require enrollment; it does not block the free first look or redirect sign-in. Absent when user is null.
x-app-account-deletion:
description: An account deletion intent (pending, processing, or failed) ends access immediately, including while financial obligations delay purge. All user sessions are revoked atomically with the intent, along with user-created API keys, share links, invitations, webhooks and inbound email addresses across every business. Better Auth lookups hide retained identities and their credentials. Validation, reset, verification and challenge errors follow the native unknown-user or unknown-token behavior, including validation order and password work. Anonymous sign-up uses the generic duplicate response without sending mail or creating credentials. OAuth cannot recreate a reserved identity. Credential inserts, updates and deletes are fenced against deletion at SQL execution time. Old session and two-factor challenge cookies are ignored on the request. Deletion never sends cookie expirations, so delayed responses cannot erase another login. Auth revocation expires sessions that anchor Plaid compensation operations and retains their rows until those operations are removed by their own lifecycle. Step-up and invite acceptance return the existing unauthenticated HTTP 401 response. Removed magic-link endpoints remain HTTP 404 for everyone. The email is reserved until purge completes. Business-only deletion retains the login and other businesses, but hides and blocks the deleting business immediately, including while billing leases drain. API keys for deleting businesses return unauthenticated HTTP 401.
x-app-step-up:
window_seconds: 600
description: 'Sensitive app actions require fresh identity verification. GET /app/api/auth/step-up returns password, code, email, passkey, google, microsoft and fresh booleans. POST /app/api/auth/step-up accepts exactly one of password, code (TOTP/backup), or emailCode. POST /app/api/auth/step-up/email sends a six-digit code to the verified email of a passwordless account, at most three times per ten minutes per user. Codes are hashed, single-use, expire after ten minutes, permit five attempts, and are bound to the initiating live session, user, email and step-up purpose. Resending invalidates the previous code. Verification is rate-limited. Passkey ceremonies require x-arkrel-step-up: 1 on both challenge issuance and verification, biometric/PIN user verification, and a server-side purpose, user and session binding that expires after five minutes. Consumption requires the exact initiating live session. Provider users call sign-in/social with a linked google provider and additionalData.stepUp=true. The server rejects idToken shortcuts and binds OAuth state to the original user/session for five minutes. Google cannot force reauthentication and is opportunistic only. It requires a back-channel auth_time between initiation and completion; token issuance time never counts. Missing, stale or future auth_time is rejected. Passwordless users can use an email code or a registered passkey when Google cannot verify freshness. Successful passkey/provider verification rotates and elevates only the returning bound session. Password, TOTP and email checks elevate the initiating session. These app-only auth surfaces are not exposed by integrations or MCP.'
x-consumer-statement-admission:
description: App uploads and inbound email share one Free statement allowance per account. Additional app uploads return 402 FREE_STATEMENT_USED. Inbound attachments wait for an upgrade. Retries recover the original admitted attachment. Developer API per-page pricing is unchanged.
x-maintenance-bridge:
description: 'During the production cutover, an isolated maintenance Worker returns HTTP 503 for every method and path, including reads and health checks, with Retry-After: 60 and Cache-Control: no-store. API, MCP, internal, and webhook paths return error.code MAINTENANCE and error.message "Arkrel is updating. Please try again in a few minutes." MCP receives this plain JSON error instead of JSON-RPC while the bridge is deployed. Other paths return HTML. HEAD returns headers without a body. Retry after the indicated delay.'
x-maintenance-mode:
description: 'MAINTENANCE_MODE=1 freezes mutations across app, REST, MCP and engine routes. Mutations and OAuth or verification callbacks return 503 with Retry-After: 60. REST errors use code MAINTENANCE_MODE. MCP tool mutations use JSON-RPC code -32004. Reads, health and status remain available. Read-time refresh, audit, counters and automatic categorization are suspended. Cron writes are skipped; inbound email invocations fail before ingestion so delivery can be retried.'
x-stripe-billing-webhook:
path: /app/api/billing/webhook
authentication: Stripe-Signature
refund_events:
- charge.refunded
- refund.updated
- refund.failed
dispute_events:
- charge.dispute.created
- charge.dispute.updated
- charge.dispute.closed
description: Refund IDs and provider statuses are persisted. Catch-Up access and unused page credits are revoked only after a succeeded refund. Pending or failed refunds retain access. The dashboard refund response has refunded=false while an automatic refund is pending or failed, with auto=true indicating that the automatic request was submitted. Disputes suspend the charged subscription, Catch-Up year, or unused page credits. A won dispute restores only the saved access and original credit expiration, unless the payment was refunded or another dispute still holds it. Lost disputes remain suspended. Subscription updates cannot bypass a hold. Independent AppSumo codes are unaffected. The owner receives an operational email alert.
x-app-catchup-payment-clock:
path: /app/api/usage
field: catchup_purchases
items:
type: object
required:
- tax_year
- purchased_at
- payment_settled_at
- expires_at
properties:
tax_year:
type: integer
purchased_at:
type: string
format: date-time
description: Purchase record time. Do not use as an unresolved payment clock fallback.
expires_at:
type: string
format: date-time
description: Access and unused pages end at this instant, 12 months after settlement. Older test rows use their purchase timestamp until settlement is reconciled.
payment_settled_at:
type: string
format: date-time
nullable: true
description: Confirmed payment charge time, or null while confirmation is unresolved. The 30-day guarantee starts at this time.