invonetwork

First-party Python server SDK for integrating INVO into partner backends. It is the server-side counterpart to the INVO JS/Web SDK: same endpoints, same field mappings, and the same webhook HMAC scheme, so both hit the same live backend interchangeably.

Status: 3.6.0 — stable, published on PyPI (pip install invonetwork). The backend it wraps is live on sandbox + production, so you can build and test against sandbox today. Recent highlights: 3.6.0 makes the device approval grant a first-class SDK surface — begin_device_approval, poll_device_approval, confirm_device_enrollment and, the step that actually moves the money, approve_with_device_code (plus the optional complete_device_approval loop); a grant that is polled to approved and left there settles nothing, which is the failure this release exists to make impossible, and TRANSACTION_NOT_PENDING now comes back as a typed status="not_pending" result with already_settled rather than a raised error you would have to string-match; 3.1.0 makes the authentication guidance explicit — passkeys are the gold standard and the SMS-PIN completion (verify_sms_transfer/verify_sms_send) is deprecated and being phased out (documentation only — no behavior change, and no runtime DeprecationWarning, so warnings-as-errors test suites are unaffected); 3.0.0 moves the Platform Commerce card leg to INVO's hosted checkout — purchase(funding_source="card") now returns a checkout_url to send the buyer to (the 2.5.x BillingAddress/client_secret surface is removed; see the CHANGELOG migration note); 2.5.0 adds Platform Commerce (ecommerce) — a platform tenant selling items funded by balance or card, with server.platform_commerce.purchase/ get_status/refund and the platform_commerce.* webhooks (the browser card-confirm step lives in the JS SDK); 2.4.0 adds Steam transfer-policy handling (is_steam_value_non_transferable/is_non_steam_value_into_steam_blocked + DestinationGame.accepts_steam_origin_value); 2.3.0 surfaces the claim-time phone-share 409 on claim_transfer/claim_currency (+ err.phone_share_last4). Full history in the CHANGELOG. Canonical partner reference: https://docs.invo.network.

Highlights

• Device approval (RFC 8628) — the console / TV / native-client passkey path, driven end to end from your server: begin, poll, answer the on-screen match code, and approve_with_device_code — the call that settles the transaction.
• Server money flows — mint player tokens, initiate cross-game sends/transfers, run the currency-purchase flow (hosted checkout + rail selector), spend game currency on items, and Platform Commerce (ecommerce: a platform tenant selling items funded by balance or card).
• Server-only reads — player balances, inbound-pending "you have X to collect", and linked wallet identities (PII, server-only).
• Webhook verification — constant-time HMAC-SHA256, replay window, multi-secret rotation.
• Resilient — automatic retries with backoff/jitter on network errors, 429 (honoring retry_after), and 5xx — for idempotent calls only.
• Zero runtime dependencies — stdlib only (urllib, hmac, json, dataclasses). Python 3.9+.
• Fully typed — ships py.typed; passes mypy --strict.

The game secret stays on your server — it authenticates every call here via the X-Game-Secret-Key header and must never reach a browser.

Passkeys are the gold standard — don't build on SMS

If you take one thing from this README: money movement should be authorized by a passkey. Do not design your verification UX around the SMS PIN.

A passkey is a WebAuthn assertion — phishing-resistant, bound to your origin, backed by the device's secure hardware. An SMS PIN is a shared secret delivered over a channel exposed to SIM swap, SS7 interception, and social engineering. They are not two equivalent ways to approve a transfer; one is materially weaker, and INVO treats it that way (hence the 24-hour money-out cooldown after a passkey recovery — that gate exists because phone-based possession can be stolen).

The WebAuthn ceremony itself runs in the browser via the JS SDK, so from this server SDK the rule shows up in how you read initiate_*:

Prerequisite: which passkey path serves your tenant. There are two places a passkey ceremony can run, decided by whether you hold a partner passkey domain:

• No partner domain (every new title): INVO runs the ceremony on its own domain — the hosted approval page — for every platform. There is no domain to verify and nothing to configure; submitting one is refused (409 PARTNER_RP_FROZEN). Your server starts a device approval grant with server.begin_device_approval(...) (the player's session token, never the game secret; below) and the browser / console / phone opens the page — then polls, and calls server.approve_with_device_code(...) when the poll says approved, which is the step that actually moves the money. The in-app browser ceremonies return 403 WEBAUTHN_NOT_ENABLED_FOR_TENANT for you — classify it with err.is_webauthn_not_enabled_for_tenant and treat it as the expected state, not a failure (the body's hosted_flow points at device_code).
• A domain verified before the freeze (existing tenants): your passkeys are bound to that RP ID for life, so the in-app ceremonies keep working. You may still change or re-verify the domain; removing it is irreversible — it can never be re-added, and the hosted flow serves the game from then on.

So the honest sequence is: start the transfer → approve on the hosted page (or in-app if you hold a domain) → you're on the gold-standard path. The SMS PIN carries only a player who cannot or will not enrol — which is exactly why the deprecation below has no removal date attached.

⚠️ verify_sms_transfer / verify_sms_send are deprecated as of 3.1.0 and a future major version will remove them. They still work exactly as before — this release changes documentation only and deliberately emits no DeprecationWarning, so pytest -W error suites keep passing. Keep the PIN path as a genuine last resort for: users who cannot enroll (unsupported device, no platform authenticator) and users who decline.

What is not discouraged. These use a one-time code, but they are the on-ramp to a passkey — not a substitute for one. Use them freely:

• recovery_begin / recovery_complete — the passkey recovery relay; restores a passkey the user lost or deleted.
• phone_share_initiate / phone_share_approve — phone-ownership consent, not transaction authorization.
• The browser-side enrollmentBegin / enrollmentVerify grant — how a user gets a first passkey.

Beyond security, this is also an economics story: SMS costs real money per message at every scale, and INVO's architecture targets passkey/in-app verification as the primary path so the platform never depends on carrier delivery. A passkey-first integration is faster for your users, cheaper to run, and won't need migrating later.

Which method fits which platform

The right approval method is a property of the client the player is sitting in front of, not of your title. Because the WebAuthn ceremony runs browser-side, this server SDK's job is to read initiate_* and let the right client half take over — pick that half per platform:

When there is no passkey path at all — a remote approval that can't run a ceremony (a guardian, a phone's existing owner) — INVO goes email first, SMS on request: a signed link to a hosted page to the oldest verified address, and a text only when there is no verified email, the email could not be delivered, or the player asks for one. SMS is the last resort, not the primary channel.

Contents

• Passkeys are the gold standard — don't build on SMS
• Install
• Get your account & game secret
• Architecture
• Before you go live
• Configuration
• Currency purchase (real money in)
• Item purchase (spend game currency)
• Platform Commerce (ecommerce)
• Player balance
• Sends & transfers
  • Legacy fallback: the SMS PIN (being retired)
• Sends and transfers, stage by stage
  • What your server must call, in order
  • Errors on the money path
  • Common mistakes
• Inbound pending & linked identities
• Consoles and TVs — approving without a browser (RFC 8628)
• Webhooks
• Resilience & observability
• Errors
• API reference
• Versioning & stability

Install

Requires Python 3.9+. The command differs slightly by OS:

# macOS / Linux
python3 -m pip install invonetwork

# Windows (PowerShell)
py -m pip install invonetwork

Recommended — inside a virtual environment:

# macOS / Linux
python3 -m venv .venv && source .venv/bin/activate && pip install invonetwork

# Windows (PowerShell)
py -m venv .venv; .venv\Scripts\Activate.ps1; pip install invonetwork

Then import:

from invonetwork import InvoServer, InvoError, verify_webhook

No third-party runtime dependencies.

Get your account & game secret (INVO console)

Sign up, create your game, and copy its game secret in the INVO console. Use the console that matches the environment you're building against:

Build and test against the dev console + sandbox first, then switch to production for launch. Each environment has its own game secret — never mix them, and keep the secret server-side only.

Architecture (this SDK is the server half)

INVO integrations split across two trust boundaries. This package is the server half; the browser half is @invonetwork/web-sdk.

┌──────────────────────────────┐         ┌──────────────────────────────┐
│  YOUR SERVER (trusted)        │         │  THE BROWSER (untrusted)      │
│  invonetwork (this package)   │  mint   │  @invonetwork/web-sdk         │
│  • holds X-Game-Secret-Key    │ ──────► │  • holds short-lived token    │
│  • mint_player_token()        │  token  │    (~15 min, game-scoped)     │
│  • initiate_send/transfer()   │         │  • enroll/approve passkeys    │
│  • create_checkout()          │         │  • confirm_receipt / claim    │
│  • purchase_currency/item()   │         │  • balances / destinations    │
│  • verify_webhook()           │         │                               │
└───────────────┬───────────────┘         └───────────────┬──────────────┘
                └──────────────► INVO BACKEND ◄────────────┘

The game secret authenticates every call here and must never reach a browser. Mint a short-lived player token server-side with mint_player_token and hand that to the browser SDK.

Player token (session mint for an existing player)

mint_player_token mints a short-lived, game-scoped session token for a player who already exists on your game — it is not a registration call. The backend looks the player up by player_email and returns a token for their existing identity (or 404 if unknown). It only needs player_email:

token = server.mint_player_token(player_email="player@example.com")

player_phone is optional here (validated as E.164 only if you pass it, and ignored by this endpoint) — an existing email-only player still mints a token.

Where phone actually matters. A player's INVO identity encodes their phone, and cross-game money routing keys off it — but that's enforced on the money calls, not the token mint:

• initiate_send / initiate_transfer require the sender's phone (E.164), and take the recipient's phone. You can send by the recipient's phone alone — they supply their email when they claim.
• An account with no phone can't receive cross-game money or take part in the account-linking consent SMS, so make sure players have a phone at creation / enrollment (in your own player system), before they transact.

Not in this SDK (by design): the browser WebAuthn ceremonies

This is the game-secret / server-side SDK. The player-token WebAuthn ceremonies — passkey enroll, approve / step-up, confirm-receipt / claim, the enrollment OTP grant, and device link — are not here, because they run in the browser (navigator.credentials) and authenticate with the player token, not the game secret. Handle them one of two ways:

• Browser: use the JS @invonetwork/web-sdk InvoClient (enrollPasskey, approveSend/approveTransfer, confirmReceipt*, enrollmentBegin/enrollmentVerify, linkDevice), or
• Proxy: relay those player-token HTTP calls through your backend (the browser still performs the actual ceremony).

Everything else — mint, initiate, verify-SMS, claim, status, guardian, phone-share, passkey-recovery relay, checkout, purchase, item purchase, balances, inbound-pending, destinations, linked-identities, and webhook verification — is in this SDK.

Passkey recovery relay (recovery_begin / recovery_complete)

The recovery calls themselves are plain OTP posts (no WebAuthn), so a Python backend can relay them. When the browser's enrollment is blocked with a 409 ENROLLMENT_REQUIRES_PROOF (the player deleted/lost their passkey — the server can't know a device-side key is gone), offer "Lost or replaced your passkey?":

# authed with the SDK player token (mint one, or relay the browser's) — NOT the game secret
server.recovery_begin(player_token=token)                 # code -> email on file first (text only if no email)
server.recovery_complete(player_token=token, code=otp)    # deactivates the stale passkey
# then the BROWSER re-runs the normal enrollPasskey() ceremony — it now succeeds

• recovery_begin sends the code email first; a text goes out only when there is no email on file. The SDK method takes no channel argument — for a "text me instead" tap, POST /api/sdk/device/recover/begin yourself with the player token and body {"channel": "sms"}. Errors: no_channel_on_file (422), rate_limited (429 — max 5 codes / 10 min).
• recovery_complete errors: ENROLLMENT_CODE_INVALID (wrong/expired, attempt-capped), RECOVERY_FAILED (500, transient).
• A player who lost the phone their INVO passkey was on can also recover from the hosted approval or claim page with no in-app call at all — see the console section below.
• Step 3 — the WebAuthn create() ceremony — can only run in the browser (JS SDK enrollPasskey()); these two calls just clear the way for it.

⚠️ 24-hour money cooldown after recovery. A recovery-enrolled passkey logs in and collects funds immediately, but money-OUT approves return 403 PASSKEY_RECOVERY_COOLDOWN for 24 hours (SIM-swap protection). Branch on err.is_passkey_recovery_cooldown, show — "For your security, transfers are paused for 24 hours after a passkey reset. You can still receive funds. Try again after err.retry_after_at." — do not retry-loop it.

Before you go live

INVO enables each flow for your tenant in the console. What to do:

• Store the game secret server-side (env var / secret manager) and expose a small endpoint that calls mint_player_token so your front-end can fetch/refresh a player token.
• Make sure players have a phone (E.164) at creation/enrollment — it's required on the money calls (initiate_send/initiate_transfer) and for cross-game receive, though not on the token mint itself (see Player token).
• Set your webhook signing secret and verify every delivery with verify_webhook — grant currency/items off webhooks, not synchronous responses.
• For currency purchase: hosted checkout works out of the box; ask INVO to enable the game/steam rails if you need them. The steam rail additionally requires the studio to register their own Steam app id + publisher Web API key and put a payment method on file — see the Steam section below.
• For sends/transfers with passkeys: nothing to configure — new titles approve on INVO's hosted page (your server starts the device approval grant; the browser half uses @invonetwork/web-sdk's approveHosted(), consoles show the QR, the mobile plugins open the system browser). Only a tenant that verified a partner domain before the freeze uses the in-app ceremony on its own origins. Wire the approve step before launch: until a sender is enrolled they fall back to the deprecated SMS-PIN path, which is the flow you don't want your users on — see Passkeys are the gold standard.
• For item purchase: nothing extra — it's a currency-balance debit.

If a flow isn't enabled yet, calls return a clear InvoError (e.g. TENANT_NOT_MIGRATED, WEBAUTHN_NOT_ENABLED_FOR_TENANT, flow_paused) — coordinate with your INVO contact to turn it on.

Configuration

import os
from invonetwork import InvoServer, Hooks

server = InvoServer(
    game_secret=os.environ["INVO_GAME_SECRET"],       # server-side only
    base_url="https://sandbox.invo.network/sandbox",  # prod: "https://invo.network"
    timeout=30,               # optional, seconds (default 30)
    max_retries=2,            # optional, default 2 (0 disables)
    retry_base_delay=0.25,    # optional backoff base, seconds
    user_agent="my-game/1.0", # optional; a sensible non-blocked UA is set by default
    hooks=Hooks(),            # optional observability (see below)
)

base_url must be https:// — the game secret travels in a request header, so plaintext is rejected. http://localhost (and loopback) is allowed for local development only.

Construct one InvoServer and reuse it. All request methods are keyword-only for clarity.

Currency purchase (real money in)

Buy game currency with real money. Authenticated by the payment rail, not a passkey.

Hosted checkout (recommended — you never touch card data)

result = server.create_checkout(
    player_email="p@example.com",
    usd_amount="20.00",                 # USD, 0 < x <= 999.99
    rail="platform",                    # optional: "platform" (default) | "game" | "steam"
    success_url="https://you/buy/ok",
    cancel_url="https://you/buy/cancel",
    metadata={"your_order_id": "ord_42"},  # echoed on the purchase.completed webhook (all rails); order_id also reconciles
)
# -> send the browser to result.checkout_url. Token TTL is result.expires_in_seconds (~900s).

The INVO-hosted page handles card entry, saved cards, and 3-D Secure. Reloading the URL after a completed payment is idempotent — it shows an already-complete success screen, not an error. Grant currency off the purchase.completed webhook, not this response.

Payment rails (neutral names)

rail selects who processes the payment. Use the neutral names; INVO enables the ones your tenant is approved for.

Steam titles: before any of this works, the studio has setup to do

Steam pays whoever owns the Steam application. Every INVO title therefore sells currency through its own Steam app, and the money lands in the studio's own Steamworks account. INVO issues the player's currency the moment Steam captures the charge, then settles with the studio separately against a payment method they keep on file.

Once per title, in the INVO dashboard:

1. Enable in-game purchases (microtransactions) for the app in Steamworks.
2. Create a publisher Web API key — a publisher group key, not a personal user key, with the app in the group. A personal key fails verification.
3. Enter the app id and that key under the title's Steam settings. INVO calls Steam to prove the pair before saving it.
4. Add a payment method on the Billing screen and accept the authorisation.

Until all four are done, purchases are refused before the player is charged: STEAM_NOT_CONFIGURED (503) or PARTNER_BILLING_NOT_SET_UP (409). That is deliberate — the alternative is taking a player's money for currency that cannot be issued.

This changed in September 2026. Versions before 3.4.0 said Steam purchases ran through an INVO-owned Steam app and there was nothing to register. That was true while one INVO application billed every title; it is not true now, and following the old text leaves the rail closed.

Steam titles: currency must be bought through Steam, and INVO sets the packs.

A title distributed on Steam may only sell currency on the steam rail — any other rail is refused with 409 STEAM_PURCHASE_LAYER_REQUIRED. This is a platform requirement, not an INVO preference, and it applies to every storefront the title ships on because a player's balance is shared across them.

You do not set Steam prices and you do not send an amount. INVO defines the pack catalogue for every game on the network. Fetch it with GET /steam/packs?steamid=..., render it, and pass the pack_id back to the purchase call.

The price is the same everywhere; the currency inside is not. Steam prices are VAT-inclusive wherever VAT is collected, so more of a fixed price goes to tax in a high-VAT country and less is left to buy currency with — the same $9.99 pack yields 69 units in the US and 58 in France. Always pass steamid when fetching the catalogue, or the amounts are quoted with no VAT deducted and your store promises more than the purchase delivers. Do not cache one catalogue for all players and do not hard-code amounts: pack prices and VAT rates are both configuration.

Omit rail to use "platform". Amounts are USD, 0 < x <= 999.99.

Steam purchases (packs, not prices)

Steam has first-class SDK methods as of 3.5.0: steam_packs(), steam_init_purchase(), steam_finalize_purchase() — all on InvoServer, all server-side. purchase_currency still rejects rail="steam" with WRONG_RAIL_ENDPOINT and points you here.

INVO owns the pack catalogue. You never send a price. Fetch the packs, show them, pass a pack_id back.

# 1. The catalogue, priced for THIS player. Always pass steamid.
result = server.steam_packs(steamid=steam_id)
# result.packs: [SteamPack(pack_id, label, price_usd, currency_amount), ...]

# 2. Start the purchase with the pack the player chose.
init = server.steam_init_purchase(
    player_email="player@example.com",
    steamid=steam_id,
    pack_id="steam_medium",                # NOT a price
    purchase_reference=my_idempotency_key, # unique per INTENDED purchase; reuse on retry
    metadata={"player_id": my_player_id},  # echoed on the purchase.completed webhook
)
# -> init.order_id, init.steam_transid, init.charged_usd, init.currency_amount

# 3a. CLIENT SESSION (default): Steam shows its in-game overlay; your game
#     client receives MicroTxnAuthorizationResponse_t and tells your backend.
# 3b. WEB SESSION: for platforms where the overlay does not render (observed:
#     Electron on macOS). Pass usersession="web" + player_ip (the PLAYER's IP,
#     never your server's) and open init.steam_checkout_url as a TOP-LEVEL
#     browser tab — it cannot be iframed.

# 4. Finalize: INVO re-checks the authorization with Steam, captures, credits.
done = server.steam_finalize_purchase(order_id=init.order_id)
# -> done.status == "success", done.new_balance, done.already_processed

The price is identical everywhere; the currency inside is not. Steam prices are VAT-inclusive wherever VAT is collected, and the storefront's revenue share comes out before currency is derived — the same $9.99 pack yields 69 units in the US and 58 in France, and a $4.99 pack yields 34, not 50. Render currency_amount verbatim; never compute currency from the sticker price.

Finalize is idempotent, and not every 409 means "try later". A replay returns already_processed=True and never credits twice. For a poller, branch on the error:

A poller that treats every 409 as "not yet" polls a dead order forever. A missed finalize is not a lost credit — INVO's reconciler independently detects the authorized transaction and credits, on a delay.

Refusals before any charge, so a player is never charged for currency that cannot be issued:

• err.is_partner_credit_unavailable — the studio's settlement float cannot cover it. Plain failure toast; do not hot-poll init (each attempt briefly re-reserves headroom), and reuse the same purchase_reference on a manual retry.

• err.is_partner_rail_suspended — terminal from the game's seat until the studio resolves it in the INVO console.

• err.is_unknown_steam_pack — re-fetch steam_packs(); the body's valid_pack_ids lists what exists.

• err.is_steam_not_configured (503) / err.is_steam_rail_not_entitled (403) — the studio's one-time Steam setup is incomplete.

• A locked Steam account is refused up front with 409 STEAM_ACCOUNT_LOCKED.

• Always pass steamid to steam_packs(). Without it the amounts are quoted with no VAT deducted — the most any pack yields — so your store promises more than the purchase delivers.

• Do not cache one catalogue for all players, and do not hard-code amounts. Pack prices and VAT rates are configuration and change without an SDK release.

• On a Steam-distributed title, a channel key may only sell on its own channel's rail. A mismatched caller is refused with 409 STEAM_PURCHASE_LAYER_REQUIRED. (The old absolute form -- "only the Steam rail, across every storefront" -- was the rule before 2026-08-31 and no longer holds for channel-keyed callers.)

• Correlation is metadata, not purchase_reference. The webhook payload deliberately omits your idempotency key; put your own player/order ids in metadata and read them back at data["metadata"] on purchase.completed -- note the Steam rail ADDS its own context keys (steam_wallet_currency, steam_country, steam_account_status, vat_rate_pct) into the echoed object, so avoid those names in yours (data["channel"] is the channel of the KEY you presented -- "steam" when you call with your Steam channel key; None on a legacy game key).

Direct rail (advanced — you tokenize the card yourself)

import uuid

purchase = server.purchase_currency(
    player_email="p@example.com",
    usd_amount="20.00",
    purchase_reference=str(uuid.uuid4()),  # idempotency key, required
    rail="platform",
    payment_method_id="pm_...",            # a tokenized payment method
    metadata={"your_order_id": "ord_42"},
)

if purchase.status == "success":
    pass  # captured; purchase.new_balance updated
elif purchase.status == "requires_action":
    # 3-D Secure: run the client action with purchase.client_secret, then:
    server.confirm_payment(payment_intent_id=purchase.payment_intent_id)
elif purchase.status == "pending_payment":
    pass  # redirect the browser to purchase.payment_url (game rail)

rail="steam" is rejected here (WRONG_RAIL_ENDPOINT) — Steam uses its own in-client flow. Reconcile with server.get_order_details(order_id=...). Most integrations should prefer hosted checkout.

Item purchase (spend game currency)

Spend the currency a player already owns to buy an in-game item. A balance debit — no real money, no payment rail, no passkey — server-side only. Amounts are in game-currency units.

import uuid

item = server.purchase_item(
    client_request_id=str(uuid.uuid4()),  # idempotency key, unique per game
    player_email="p@example.com",
    player_name="P",
    item_id="sword_001",
    item_name="Legendary Sword",
    item_quantity=1,                       # integer 1..1000
    unit_price="100.00",                   # > 0 and <= 999999.99
    total_price="100.00",                  # must equal unit_price * item_quantity (+/-0.01)
    # optional: player_phone, item_description, item_category
)
# item.status == "success"; item.new_balance / item.previous_balance / item.currency_name
# item.transaction_id / item.order_id; item.financial_breakdown

• Grant the item off the item.purchased webhook, not just this response. INVO debits currency and records the purchase; your game owns the item catalog and grants the item.
• Idempotent on client_request_id — a duplicate raises 409 (err.is_duplicate_request).
• Insufficient balance raises 400 (err.is_insufficient_balance; required_amount + current_balance on err.body).
• Client-side validation (missing fields, quantity outside 1..1000, bad price, total mismatch) raises INVALID_INPUT before any network call.

Companion reads: get_item_purchase_history(player_email=..., limit=?, offset=?) and get_item_order_details(order_id | transaction_id | client_request_id) (pass exactly one id — use client_request_id for recovery: "did this purchase complete?"). To walk the full history, iterate — it pages automatically:

for row in server.iterate_item_purchase_history(player_email="p@example.com"):
    ...

Platform Commerce (ecommerce)

This is not item purchase. Item purchase is a game tenant spending a player's existing game currency on an in-game item — always a balance debit, never a card, no refunds. Platform Commerce is a platform tenant (a non-game app: vertical video, creator merch, marketplace) running a storefront: the buyer pays with INVO balance or a real card (new money), INVO is merchant of record, and refunds exist. Only platform tenants may call it — a game tenant gets 403 (err.is_not_platform_tenant).

The funding source is resolved server-side under lock — the client can request balance or card, but the backend verifies the real balance before value moves. This SDK is the server half: it creates purchases and refunds. On the card leg, INVO hosts the entire checkout (card fields, Apple Pay / Google Pay, billing-address collection, 3-D Secure) — the server call returns a checkout_url and your app just sends the buyer there. There is no billing address in the request and no client payment code to write.

import uuid

# Balance leg — settles synchronously
r = server.platform_commerce.purchase(
    client_request_id=str(uuid.uuid4()),   # idempotency key, unique per tenant
    funding_source="balance",
    player_email="user@example.com",
    player_name="Ada",
    item_id="sticker_pack_01",
    item_name="Sticker Pack",
    item_quantity=1,                        # integer 1..1000
    unit_price="5.00",                      # BALANCE leg: the tenant's network-currency amount
    total_price="5.00",                     # must equal unit_price * item_quantity (+/-0.01)
)
# r.status == "success"; r.new_balance / r.currency_name / r.order_id
# r.financial_breakdown  # INVO fee: 3.5% flat

# Card leg — returns a hosted-checkout session; NOT yet paid. total_price is USD ($0.50–$999.99).
r = server.platform_commerce.purchase(
    client_request_id=str(uuid.uuid4()),
    funding_source="card",
    player_email="user@example.com",
    player_name="Ada",
    item_id="sticker_pack_01",
    item_name="Sticker Pack",
    item_quantity=1,
    unit_price="5.00",                      # CARD leg: USD
    total_price="5.00",
    success_url="https://app.example/thanks",  # optional: where the buyer lands after paying
    cancel_url="https://app.example/cart",     # optional: where the buyer lands on cancel
    metadata={"cart_id": "c_9"},               # optional: echoed back on the webhook
)
# r.status == "requires_payment"
# r.checkout_url  → send the buyer here (redirect, or the JS SDK's mountCheckout embed)
# r.session_id / r.expires_at (unix seconds)

# Status + refunds
s = server.platform_commerce.get_status(r.order_id)
# s.status: "completed" (balance now; card after the webhook) | "pending_payment" | "refunded"

ref = server.platform_commerce.refund(order_id=r.order_id, reason="customer request")
# or refund(client_request_id=...). Pass EXACTLY ONE id.
# INVO retains its fee (ref.fee_retained is True); the customer is made whole minus that fee.
# A second refund of the same order raises 409 (err.is_already_refunded) — treat as already done.

• Fulfill card orders on the platform_commerce.purchased webhook, never on the client return — the sale is real only once the payment settles on INVO's hosted page.
• Idempotency on client_request_id: a duplicate BALANCE purchase raises 409 (err.is_duplicate_request); a duplicate CARD purchase replays the same checkout session (r.idempotent_replay is True, same URL — never a second charge), so a lost card-leg response is safely recovered by retrying with the same client_request_id.
• INVO fee: 3.5% flat (balance) · 3.5% + $0.30 (card). Too-small amounts raise err.is_amount_below_minimum / err.is_below_card_minimum (card < $0.50); a card total over $999.99 raises err.is_above_card_maximum.
• Client-side validation (missing fields, bad funding_source, quantity outside 1..1000, total mismatch, card USD bounds, non-E.164 player_phone) raises INVALID_INPUT before any network call.

Player balance

result = server.get_player_balance(player_email="p@example.com")
# Lookup is by EMAIL only — there is no by-id balance route (player_id is a per-game internal
# id). For a client-side read, use the browser InvoClient.getBalance() (identity from the token).
for b in result.balances:
    print(b.currency_name, b.available_balance, b.total_balance)

Sends & transfers

Move already-owned game currency from one player to another, authorized by the sender's passkey — the gold standard.

There are exactly two approval paths, and both are passkeys:

• In-app passkey — wherever WebAuthn can run in the client: the browser, via the JS SDK (approveSend / approveTransfer), or INVO's hosted approval page (approveHosted).
• Device approval (QR) — wherever it cannot: consoles, TVs, native Steam and desktop clients. This SDK drives that one end to end — the player approves on their phone and your server then makes the call that moves the money. See Consoles and TVs and Sends and transfers, stage by stage.

A legacy SMS-PIN fallback still exists for a sender who can use neither — it is being retired and new integrations should not build on it.

The server initiates:

import uuid

t = server.initiate_transfer(
    client_request_id=str(uuid.uuid4()),
    source_player_name="P",
    source_player_email="p@example.com",
    source_player_phone="+15555550100",
    target_player_email="q@example.com",
    target_player_phone="+15555550111",
    target_game_id=123456,
    amount="50",
)
# initiate_send uses sender_*/receiver_* + receiving_game_id instead.

# Check guardian_approval FIRST — the guardian path takes precedence.
if t.guardian_approval:
    ...  # minor/guardian path (HTTP 202): pending approval, do NOT show a PIN UI
elif t.verification_method == "in_app":
    ...  # sender HAS a passkey -> approve in the browser (JS SDK). The good path.
else:
    ...  # sender has NO passkey yet -> enrol one (browser enrollPasskey() / approveHosted()),
    ...  # or run the QR device-approval grant on a console. See "stage by stage" below.

On the guardian path verification_method is None (even though the raw 202 body also carries "sms") so guardian_approval wins — but branch on it first to be safe.

Both games must be Live. initiate_send/initiate_transfer raise 403 if either side is still in testing, and the two cases are deliberately separate because the fix differs: err.is_source_game_not_live is your game (self-serve — switch it to Live in the console under Game Settings > Status), err.is_target_game_not_live is the destination game, usually owned by another developer, so you can't flip it — show err.message and steer the player elsewhere via get_destinations. err.game_status carries the current state. Don't merge these into one "not live" branch; it produces a "go fix it" action that leads nowhere half the time.

• The one message that is still a text: the claim link to a phone-only recipient. A send can be addressed to someone INVO has never met — the sender knows their phone number and nothing else, so there is no email to write to and no session to notify. That is the only remaining place INVO sends an SMS by default; everything else that needs a human (guardian consent, enrolment backup, phone-share) goes by email first. The claim text a send's receiver gets now ends with a link to a hosted INVO page ("Or collect here: …"). They see who sent what in which game, enter one email (the same thing in-game claim already requires), confirm it with an emailed 6-digit code, and their phone's fingerprint / face / screen lock creates an INVO passkey — a recipient who already has one just confirms with it. The currency lands in their balance for the receiving game; the page shows the new balance and tells them to sign in to the game with that email + phone. The in-game claim is unchanged — point a player at the link when they don't have the game open, at in-game claim when they do; both are the same claim with one lifetime. transfer.received fires the same either way, and you change nothing.

• Resend the claim SMS — a plain REST call, not an SDK method. Authed with the sender's player token (not the game secret), no body. It re-sends the identical message (same code, same link); nothing new is minted.

  r = requests.post(
      f"{INVO_BASE}/api/sdk/send/{transaction_id}/resend-claim",
      headers={"Authorization": f"Bearer {sender_player_token}"},
  )
  # 200 {"status": "resent", "transaction_id": ..., "retry_after": 30}
  # 429 RESEND_COOLDOWN (+ retry_after) — 30-second cooldown; max 10 per transaction per hour
  # 400 CLAIM_EXPIRED / NOT_CLAIMABLE — the send can no longer be claimed; start a new send
  # 403 — the token is not the send's sender
  

• Uncollected after 24 hours → refunded, and both parties are told. The refund to the sender is as before; what's new is that the sender gets an email and the recipient a single text saying the currency was returned, instead of silence. transfer.claim_expired + transfer.refunded remain the source of truth — reconcile off those, not the messages.

• Guardian approval goes by email first. When a minor's initiate returns the 202, the guardian gets a signed, single-use link to a hosted INVO page with Approve / Decline — not a text. A reply-YES text goes out only when the guardian has no verified email or the email could not be delivered. The raw 202 guardian_approval block (t.raw) now carries consent_channel ("email" | "sms") — use it for your waiting copy — and resend_endpoint. Poll get_guardian_approval_status exactly as before; the approval object's consent_channel says how it went out and decision_source is email_link for a page decision.

• "Text my parent instead" — a plain REST call, not an SDK method. Authed with the minor's player token (the session that started the transaction), once per approval:

  r = requests.post(
      f"{INVO_BASE}/api/sdk/approvals/guardian/{approval_id}/resend",
      headers={"Authorization": f"Bearer {initiator_player_token}"},
      json={"channel": "sms"},   # the only accepted value
  )
  # 200 {"status": "sent", "channel": "sms"}
  # 400 INVALID_INPUT   — channel was not "sms"
  # 403 NOT_INITIATOR   — the token is not the minor who started it
  # 404 APPROVAL_NOT_FOUND (opaque)
  # 409 CHANNEL_WAS_SMS — it already went by text; hide the button when consent_channel == "sms"
  # 409 ALREADY_RESENT  — one text per approval
  # 410 APPROVAL_GONE   — already decided or expired
  # 503 DELIVERY_FAILED — retryable; does not use up the one allowance
  

  The emailed link keeps working alongside the text; whichever answers first wins.

• Phone-share and recipient-identity consents use the same hosted page. On err.is_phone_share_approval_required the phone's existing owner gets an emailed link (Allow / Decline) when INVO has itself proven an address for them — at most three holders, one email per request — and the 409 body carries consent_channel: "email"; the OTP text goes out instead only when no proven holder exists (consent_channel: "sms"). The RECIPIENT_IDENTITY_PENDING hold emails the phone owner's oldest verified address the same way. Nothing changes for you: same 409 / 202 codes, same poll-and-retry, and phone_share_approve (typed code) and the in-app approve still work.

Legacy fallback: the SMS PIN (being retired)

This is the legacy path. It is being retired, and new integrations should not build on it. It is kept working only for a sender who genuinely cannot approve with a passkey: an unsupported device with no platform authenticator, or a player who declines both the in-app ceremony and the QR flow.

When a sender has no enrolled approver, initiate_send / initiate_transfer return verification_method == "sms" and a one-time PIN is texted. The completion calls are server.verify_sms_transfer(transaction_id, pin) / verify_sms_send(...) — deprecated since 3.1.0, still fully functional, and with no runtime DeprecationWarning so warnings-as-errors suites are unaffected. They will be removed at some future major; no date is being promised.

Why it is going away: a PIN in a text is a shared secret on a channel exposed to SIM swap, SS7 interception and social engineering, it costs real money per message at every scale, and it depends on carrier delivery INVO does not control. A passkey — in the client where WebAuthn can run, on the player's phone by QR where it cannot — is stronger, faster and free. Read verification_method == "sms" as "this sender has no passkey yet" and offer them one; show a PIN pad only when there is no other way to serve that specific player.

Sends and transfers, stage by stage

One reference for the whole money path — both flows (send and transfer) and both sides (the sender who authorises, the recipient who collects). It exists because the stages are easy to read as three when there are four, and the fourth one is the money.

Companion page, same title, on the docs site: Sends and transfers, stage by stage — the same walkthrough with the raw REST calls, for stacks that do not use this package.

The four stages.

1. Initiate (your server). Nothing has moved; a transaction now exists, pending the sender's approval.
2. Prove who is approving. A passkey ceremony in the browser (the JS SDK), on INVO's hosted page, or — where no browser exists — a device approval grant on the player's phone, driven from here.
3. Settle (your server, for the grant path). A proof is not a payment. The approval says who; a separate call moves the value and runs every gate (guardian, risk, recovery cooldown). In the browser the approve call does both at once, which is why the split only becomes visible on the grant path.
4. Collect. The recipient confirms receipt (their own passkey, or the grant), or redeems the claim code. Reconcile off webhooks, never off stage 3's response alone.

What your server must call, in order

Steps 3c, 4 and 5b take the player token explicitly — on a console the game server holds the token it minted, and the game secret is never sent on any of them.

The four device-approval flows

flow picks which endpoint stage 4 settles against. Begin with one flow and settle with the same one: a grant approved for transfer settles nothing else.

Errors on the money path

Also reachable from a receipt settle (send_receipt / transfer_receipt) — a console recipient built only from the table above ships with no fallback for these:

⚠️ A settle FAILURE is ambiguous about the money. INVO commits the transfer and then builds its 200 by reading the transaction back — so a fault raised after that commit answers 400 *_FAILED for money that already moved. An err.status == 0 (timeout, dropped connection) is ambiguous for exactly the same reason, and this call is deliberately never auto-retried. On either: read get_transfer_status / get_send_status, or wait for the transfer.claim_pending webhook, before you tell the player it failed — and never auto-start a replacement transaction.

"Already settled" is not a refusal

The integrator question this section was written for: did the guard stop me, or is her gold stuck? You must be able to answer that without matching on message text.

approve_with_device_code does not raise for TRANSACTION_NOT_PENDING. It returns:

r = server.approve_with_device_code(
    player_token=player_token,
    transaction_id=transaction_id,
    flow="transfer",
    device_code=device_code,
)
if r.status == "not_pending":
    if r.already_settled:
        # This step already happened — typically your first attempt landed and the
        # response was lost. r.current_status says where it is now
        # (pending_claim, completed, ...).
        ...
    else:
        # NOT settled. Either dead (expired / failed / refunded), or — on a receipt flow —
        # not there YET: the sender has not approved, or their approval is held at a
        # guardian. It is also what an unrecognised status returns, deliberately.
        # Go and read it; do not guess and do not start a replacement transaction:
        t = server.get_transfer_status(transaction_id)
        ...
else:
    # status "approved" (+ next="pending_claim", r.claim_code on a transfer)
    # or "completed" (+ r.amount_received) on the receipt flows
    # or a 202 hold — r.hold_reason names it.
    ...

Every other refusal is a normal raised InvoError with a typed property, so a try/except around this call still means "something is wrong". Nothing in your code has to read an English sentence to tell the two apart.

already_settled fails closed. It is True only for a status INVO knows to be past this flow's step — and the two receipt flows have a different set from the two sender flows, because the four endpoints gate on different statuses. Anything else, including a status this SDK version has never seen, answers False with current_status set. Cautious and wrong costs you a status lookup; confident and wrong credits goods for money that never arrived.

(Older INVO deployments answered this case with only a message — "Transaction not pending PIN verification…", "…cannot be confirmed…" — and no machine code. The SDK matches those shapes too, so already_settled is correct against an older backend without any string-matching in your code. That fallback requires HTTP 400, one of those two phrases, and a readable (current status: …) / (status: …) clause: "cannot be confirmed" is ordinary English, so a phrase match alone is never allowed to assert that money moved.)

Common mistakes

1. Polling to approved and stopping. The one that costs real money. A device approval grant is a factor, not a payment: approved means the player proved who they are, and the transaction is still sitting at pending_pin_verification. You must call approve_with_device_code. Nothing else in this SDK settles it, and no webhook does it for you — device_approval.approved is the cue to call it. Left alone, the transaction expires on its own window and is refunded.

2. Treating "already settled" as a failure. A repeat call after a crash or a lost response answers TRANSACTION_NOT_PENDING. That is your earlier attempt succeeding, not a refusal. Branch on r.already_settled — never on the message.

3. Sending the game secret on the grant calls. Every device-approval call authenticates with the player token. A game secret in a console build is a game secret in a disassembler.

4. Beginning a second grant instead of settling the first. If begin answers 409 DEVICE_APPROVAL_ALREADY_PENDING, the player has already approved; you need stage 4, not another QR.

5. Polling faster than interval. You get slow_down; add 5 seconds and resume.

6. Using the wrong flow at stage 4. The sender's grant (transfer) cannot settle the recipient's step (transfer_receipt) — it answers DEVICE_APPROVAL_NOT_APPROVED, which reads like a broken code and is really a wrong flow.

7. Granting value off stage 3's HTTP response. Reconcile off webhooks.

Inbound pending & linked identities

"You have X to collect" (server, game-secret): the player's incoming, unclaimed sends/transfers — including value sent from other games to a player on your platform.

pending = server.get_inbound_pending(player_email="p@example.com")  # or player_phone=...
for row in pending.inbound_pending:
    # Match row.to_phone to the logged-in player. row.to_identity_id is None when the phone
    # maps to more than one of your players — don't require it.
    print(row.transaction_id, row.net_amount, row.to_phone, row.source_game)

• Lists only pending/unclaimed inbound; once claimed it drops off. row.source_game is where it came from (another game/platform). Pairs with the transfer.claim_pending webhook (the webhook is the wake-up; this is the list).
• This is the server/platform view (game-secret). The browser player-token equivalent lives in the JS SDK as client.getPendingCollect() (there, incoming rows are kind="receiving_confirm").

Linked wallet identities (server-only — returns PII):

ident = server.get_linked_identities(player_email="p@example.com")  # phone wins if both given
if ident.not_found:
    ...  # no in-game match (backend 404) — treat as "no linked identities", not an error
else:
    print(ident.primary_email, ident.is_minor, [e.email for e in ident.emails])

⚠️ Returns first-party PII (emails/phones) — never expose this to the browser.

Consoles and TVs — approving without a browser (RFC 8628)

A console has no browser, so it cannot run a passkey ceremony at all. Device Approval moves the ceremony to a device that can: your game shows a short code, the player approves on their phone, your server polls until it hears back — and then your server makes the call that actually moves the money. It is the OAuth 2.0 Device Authorization Grant, RFC 8628 — the same flow every console and TV app uses — so a standard client library understands the responses.

The grant is a factor, not a payment. Reaching approved proves who approved and settles nothing. The transaction is still waiting. approve_with_device_code is the money step, and without it the transaction expires and is refunded. This is the single most common integration failure on this flow — see Sends and transfers, stage by stage.

Five SDK methods, since 3.6.0 (begin_device_approval, poll_device_approval, confirm_device_enrollment, approve_with_device_code, and the optional complete_device_approval loop). Earlier versions documented this as raw REST only, which is how a grant could be polled to approved and then dropped. The REST endpoints are unchanged and still usable directly.

Every one of these calls authenticates with the player token, never the game secret — on a console it is your server that holds the token it minted for the player.

The endpoints underneath, for a stack that is not using this package. All four take Authorization: Bearer <player token>; none takes the game secret:

Each settle call takes the body {"device_code": "..."} and the player's bearer token — the same token that began the grant.

Available to every title, with nothing to configure. The constraint is a property of the device, not your game: a title with a website and a console build uses the normal ceremony in the browser and this one on the console.

Native Steam and desktop game clients (Mac & Windows) belong here too. Even though the operating system supports passkeys, an embedded game client can't invoke the platform authenticator from inside the engine — so a native desktop title uses this QR flow exactly as a console does. It is the client that forces QR here, not the OS.

import time
from invonetwork import InvoError

player_token = session["invo_player_token"]   # minted by mint_player_token()

# 1. BEGIN — start an approval for ONE transaction.
#    flow:    "transfer" | "send" | "send_receipt" | "transfer_receipt"
#    channel: "qr" (default — console / TV / native desktop) or "app_browser"
#             (the Unity / Unreal plugin opens the system browser; INVO derives the
#             return scheme invo-sdk-<game_id> itself, never caller-supplied).
#             "popup" is BROWSER-only — the page must call begin itself so its https
#             Origin becomes the popup's opener; that is the JS SDK's approveHosted().
#             This SDK refuses it before the network.
grant = server.begin_device_approval(
    player_token=player_token,
    transaction_id=transaction_id,
    flow="transfer",
)
# grant.device_code / user_code / verification_uri / verification_uri_complete
#      / expires_in / interval / channel

# 2. SHOW — render grant.verification_uri_complete as a QR and print grant.user_code
#    under it. The QR already carries the code, so most players never type.
#    Keep grant.device_code on your server: it is a bearer credential.

# 3. POLL — never faster than grant.interval seconds.
interval = grant.interval
while True:
    time.sleep(interval)
    poll = server.poll_device_approval(
        player_token=player_token, device_code=grant.device_code
    )

    if poll.status == "authorization_pending":
        # A first-time phone may be waiting on YOUR screen — see `enrollment` below.
        if poll.enrollment:
            show_match_prompt(poll.enrollment)
        if poll.interval:
            interval = poll.interval
        continue
    if poll.status == "slow_down":        # RFC 8628 §3.5
        interval += 5
        continue
    if poll.status == "access_denied":
        return player_declined()
    if poll.status == "expired_token":
        return start_over()               # begin again

    # 4. APPROVED -> **CALL THE MONEY STEP.** Nothing has moved until this returns.
    settle = server.approve_with_device_code(
        player_token=player_token,
        transaction_id=transaction_id,
        flow="transfer",                  # the SAME flow you began with
        device_code=grant.device_code,
    )

    if settle.status == "not_pending":
        # Your earlier attempt already landed (settle.already_settled) or the
        # transaction is dead (expired / refunded). Read settle.current_status.
        # NOT a refusal.
        ...
    else:
        # "approved" + next="pending_claim" (+ settle.claim_code on a transfer), or
        # "completed" + settle.amount_received on the receipt flows, or a 202 hold
        # (settle.hold_reason).
        ...
    break

Or let the SDK run stages 3 and 4 — complete_device_approval polls on interval, adds 5 s on slow_down, surfaces the enrolment prompt, and on approved calls approve_with_device_code for you:

def on_enrollment(enrollment, respond):
    if enrollment is None:
        return hide_match_prompt()
    show_match_prompt(enrollment)        # "Set up INVO on <device_label>? Code <match_code>"
    respond("approve" if player_said_yes() else "deny")

done = server.complete_device_approval(
    player_token=player_token,
    transaction_id=transaction_id,
    flow="transfer",
    device_code=grant.device_code,
    interval=grant.interval,
    on_enrollment=on_enrollment,
    should_stop=lambda: shutting_down,   # optional
)
# done.status: "approved" (done.settlement carries the result) | "denied" | "expired" | "stopped"
#   ("stopped" is your should_stop callback; the JS SDK calls its equivalent "aborted",
#    because there it is an AbortSignal. Same situation, different mechanism.)

Convenience only. It blocks for as long as the player takes, so it suits a worker or a job, not a request handler with a short timeout. And because it moves money from inside a loop, record the attempt in your own store before you call it — if the process dies between the approve call and its response, the only way to learn whether the money moved is a record you wrote first, plus get_transfer_status / get_send_status. Servers that cannot block should use the three calls directly and keep device_code in their own store.

• One approval authorises one transaction. Unlike the plain RFC 8628 grant, which authorises a client, an INVO device code is bound to the transaction and the flow you named. It cannot approve anything else.

• When the code is spent: at the money step, not at approval. The device_code stops being usable when approve_with_device_code settles the transaction — because the transaction is then no longer pending, and that is what the settle call checks under a row lock. Polling to approved consumes nothing; a second approve_with_device_code with the same code answers TRANSACTION_NOT_PENDING (returned as status="not_pending", not raised) — which, on a sender flow, is how you confirm that a settle whose response you lost did in fact land.

• user_code is meant to be seen; device_code is not. Poll from your backend, never from the game client, and never log or render the device code.

• One live code per transaction and flow. Calling begin again while an earlier grant is merely pending supersedes it — you get a fresh code and a page still open on the old one is told to scan again (the right answer when a player retries after a crash or an abandoned phone ceremony). If the earlier grant was already approved, the begin is refused with 409 DEVICE_APPROVAL_ALREADY_PENDING + expires_at (err.is_device_approval_already_pending, err.device_approval_expires_at): the player has approved and you are missing the money step, not a new code. Once that approved grant's own window passes, begin works again for the same transaction. The sender's grant and the recipient's are scoped separately, so transfer and transfer_receipt never block each other.

• Starting an approval extends the transaction's window. For transfer and send, begin pushes the transaction's own approval window past the code's expiry, so an approval at minute nine can't land on a transfer that has already expired and been refunded.

• Subscribe to device_approval.approved to learn the moment an approval lands instead of waiting for the next poll — then call approve_with_device_code. The webhook is a latency improvement, not the money step and not a replacement for it; the transfer.* lifecycle events that follow are what you reconcile against.

• First-time approvers enrol on the spot — your only job is one Yes/No prompt. The approval page uses a passkey registered with INVO, on INVO's own domain; a passkey the player set up on your domain is a different credential and cannot be used there — that is how passkeys work. So the first scan enrols one, and the proof it's really them is your game screen: the phone says "Confirm on your game screen", your game shows "Set up INVO on this phone: iPhone?", the player presses Yes on the console, and their phone's fingerprint / face / screen lock creates the passkey and approves. No code to type, no SMS, ever. Every later scan is scan, biometric, done. Same begin/poll, same device_approval.approved webhook, same method value device_grant_webauthn, same approve call with the device_code afterwards.

• Poll: the enrollment object. While a phone is waiting on your screen, the authorization_pending body gains it (absent when no phone asked — an already-enrolled phone never triggers it):

  {"error": "authorization_pending",
   "enrollment": {"state": "awaiting_screen", "device_label": "iPhone",
                  "match_code": "48-27", "recovery": false,
                  "requested_at": "2026-09-03T18:04:12+00:00"}}
  

  Show the prompt while state is awaiting_screen — "Set up INVO on iPhone? Code 48-27. Say Yes only if the phone you just scanned shows this code." — and remove it the moment it changes (confirmed / denied) or the object disappears. The player confirms a match, so draw match_code large. device_label is from a fixed list (iPhone, iPad, Android phone, Android device, Mac, Windows PC, Chromebook, Linux device, phone), derived from the scanning phone's browser — a hint, not an identity. Lead the copy with "scanned just now". recovery: true means the phone declared it is replacing a lost INVO passkey — word the prompt "Replace your INVO passkey with this iPhone? Code 48-27" and answer through the same confirm-enrollment (see Recovering a lost passkey below).

• Answer it: confirm_device_enrollment — the same player token as begin/poll (POST /api/sdk/approvals/device/confirm-enrollment underneath).

  r = server.confirm_device_enrollment(
      player_token=player_token,
      device_code=grant.device_code,
      decision="approve",              # or "deny"
  )
  # r.status: "confirmed" | "denied"
  # raises 409 DEVICE_APPROVAL_ENROLLMENT_ALREADY_DECIDED {"decided", "via"}
  #     -> the backup email answered first; take the prompt down quietly
  # raises 409 DEVICE_APPROVAL_NO_ENROLLMENT_PENDING -> no phone has asked
  # raises 400 invalid_grant / expired_token -> the grant is gone; the next poll says so
  

  Inside complete_device_approval the respond callable does this for you and swallows exactly those four "the prompt is moot" refusals, because a prompt nobody answered in time must not break the loop.

  Deny ends the grant: poll returns access_denied, and the phone says it was declined on the game screen.

• Backup email. At the scan, an email goes to the address on file — "confirm on your game screen; can't see the prompt? tap here; wasn't you? tap here; ignore this if you already confirmed". Opened after the screen answered, it says "you already confirmed this". A backup for a missed prompt, not the proof channel; nothing for you to do.

• Mobile (channel: "app_browser"): enrolment is auto-confirmed, no prompt. The Unity / Unreal plugins open the same page in the system browser and return on invo-sdk-<game_id>://done, which carries nothing. Because the phone running the page is the device that opened it, the game's prompt is unreachable and a match code proves nothing — so INVO confirms the enrolment itself: the poll's enrollment object appears already confirmed and goes straight to approved; the backup email still goes out as the "if this wasn't you" alert. Start polling on interval the moment begin returns — the return is only a hint to poll sooner.

• Recovering a lost passkey — from the page, nothing new on your side. A player whose only INVO passkey was on a phone they no longer have taps "Recover my passkey by email" on the page. Recovery needs both halves of the identity: your game screen confirms the phone (the recovery: true prompt above — auto-confirmed on mobile), and only then a single-use recovery link is emailed to the address on file (never an address the page supplies; one recovery per identity per 24 h). Opened on the same phone, that tab confirms and the page carries on: INVO deactivates the old INVO passkeys, notifies the owner on every channel, enqueues an identity.passkey_reset webhook to your game, and the phone enrols a fresh passkey and settles the approval — your poll reports approved and device_approval.approved fires as usual. The hosted pages are browser-only by design; there is no server-side call to make and nothing for this SDK to wrap.

  ⚠️ Money out is paused for 24 hours after a recovery. The approval settles, but the approve call for a transfer / send is refused with 403 PASSKEY_RECOVERY_COOLDOWN (err.is_passkey_recovery_cooldown, err.retry_after_at). The page tells the player "recovered; start this transfer again after the hold" — expect a fresh initiate later, not a retry. Receiving and collecting are unaffected.

• An identity with a method elsewhere must vouch for the new one. If the identity already has an active approver — a passkey on a partner domain, the INVO app's device key — the page refuses to enrol beside it with 409 ENROLLMENT_REQUIRES_PROOF and tells the player to add this phone from where that method lives (device link), then scan again. An identity with no method enrols freely.

• The code box on the page is the RFC 8628 user_code, there for a TV with no camera. When the QR pre-fills it, the box isn't shown. It is not an SMS field.

• Currency purchases on a console go through that console's store, as the platform holder requires. This flow is for approving transfers and sends.

Webhooks

Synchronous responses are for UX; reconcile and grant value off webhooks. They're HMAC-signed; dedupe on idempotency_key (stable across retries/replays).

verify_webhook does constant-time HMAC-SHA256 over f"{t}.{raw_body}", enforces a 5-minute replay window, and accepts a list of secrets during rotation. Pass the raw request bytes (never a re-parsed object).

Flask

from flask import Flask, request, Response
from invonetwork import verify_webhook, InvoError

app = Flask(__name__)
seen = set()  # replace with a durable store

@app.post("/invo/webhooks")
def invo_webhooks():
    try:
        event = verify_webhook(
            request.get_data(),                        # raw bytes — do NOT use request.json
            request.headers.get("X-Invo-Signature"),
            os.environ["INVO_WEBHOOK_SECRET"],         # or [old_secret, new_secret] during rotation
        )
    except InvoError as e:
        return Response(e.code or "invalid_signature", status=400)

    if event.idempotency_key in seen:
        return Response(status=200)                     # already processed
    seen.add(event.idempotency_key)

    if event.event_type == "purchase.completed":
        grant_currency(event.data)                      # event.data is a dict
    elif event.event_type == "item.purchased":
        grant_item(event.data)
    # transfer.*, payout.status_changed, ...

    return Response(status=200)                          # 2xx fast; offload slow work

FastAPI

from fastapi import FastAPI, Request, Response
from invonetwork import verify_webhook, InvoError

app = FastAPI()

@app.post("/invo/webhooks")
async def invo_webhooks(request: Request):
    raw = await request.body()  # raw bytes
    try:
        event = verify_webhook(
            raw,
            request.headers.get("x-invo-signature"),
            os.environ["INVO_WEBHOOK_SECRET"],
        )
    except InvoError as e:
        return Response(e.code or "invalid_signature", status_code=400)

    # de-dupe on event.idempotency_key, then grant value.
    handle(event)
    return Response(status_code=200)  # raise / return 5xx to make INVO retry

verify_webhook raises InvoError (all status == 0) with one of these codes on failure: WEBHOOK_SIGNATURE_MISSING, WEBHOOK_SECRET_MISSING, WEBHOOK_TIMESTAMP_EXPIRED, WEBHOOK_SIGNATURE_INVALID, WEBHOOK_MALFORMED. Return a 4xx on those; return a 5xx from your own handler if you want INVO to retry.

Key event types

Resilience & observability

• Automatic retries. Transient failures — network errors/timeouts, 429 (honoring retry_after, capped at 20s), and 5xx — are retried with exponential backoff + jitter. Configure with max_retries (default 2, 0 disables) and retry_base_delay. Mutating calls carry idempotency keys, so retries are safe; non-idempotent calls (e.g. hosted checkout creation) are never auto-retried.
• Hooks. Best-effort tracing/metrics (a throwing hook never breaks a request):

from invonetwork import Hooks

server = InvoServer(
    game_secret=..., base_url=...,
    hooks=Hooks(
        on_request=lambda i: log(i.method, i.url, i.attempt),
        on_response=lambda i: metric(i.status, i.duration_ms, i.request_id),
        on_error=lambda i: log(i.error.status, i.will_retry),
    ),
)

Hook payloads include the request url, which for some calls embeds a player email. The game secret is a header and is never passed to hooks — redact url if you log payloads.

• Request ids. InvoError.request_id carries the backend request id — quote it in support tickets.

Errors

Every failure raises InvoError with:

• .code — stable machine code when present (some txn-state errors have none — branch on .message)
• .status — HTTP status (0 for client-side validation and network errors)
• .message — human-readable
• .body — the raw parsed response
• .request_id — backend request id, when present

.status == 0 means "no HTTP response" — and nothing else

This is the single most important thing to get right when handling INVO errors:

The SDK never collapses a non-2xx into a transport error. A 403 arrives as status == 403 with .code and .message intact; only a genuine absence of a response produces status == 0. Covered by regression tests in both SDKs.

The failure mode to avoid is in your exception handler:

# ✗ Loses everything the API told you. A 403 with a precise, actionable refusal
#   renders as a network outage, and the developer debugs a 500 that never happened.
try:
    server.initiate_transfer(**payload)
except Exception:
    show_toast("Could not reach the server")

# ✓ Distinguish "no response" from "answered with a refusal".
try:
    server.initiate_transfer(**payload)
except InvoError as e:
    if e.status == 0:
        show_toast("Couldn't reach INVO — check your connection.")
    else:
        show_toast(e.message)      # the API already wrote the actionable text

This is not hypothetical. A developer once spent a debugging session hunting a 500 that didn't exist — the logs showed two clean 403s carrying exact remediation steps, and a catch-all in the integration layer had rendered them as "could not reach the server."

409 is usually "not ready" or "already done", not a failure

INVO uses 409 for retryable not-ready states, not errors: a duplicate client_request_id, an already-refunded order, an unverified domain. Branch on them (.is_duplicate_request, .is_already_refunded, .is_phone_share_approval_required, .is_phone_share_already_approved) and treat them as "retry" or "already done" — surfacing them as red error states misrepresents what happened.

Classifiers:

from invonetwork import InvoError

try:
    server.purchase_item(...)
except InvoError as e:
    if e.is_insufficient_balance:
        show_top_up(e.body)  # {required_amount, current_balance}
    else:
        raise

API reference

InvoServer

Construct: InvoServer(game_secret, base_url, *, timeout=30, max_retries=2, retry_base_delay=0.25, user_agent=..., hooks=None, http=None)

Module-level

Every result keeps the full backend body on .raw for fields not surfaced explicitly.

Versioning & stability

Since 1.0.0 the public API is stable: it follows semver, and no breaking change ships without a major version bump + a migration note (two majors to date: 2.0.0 required player_phone at the token mint, which 2.2.0 later relaxed back to optional, and 3.0.0 moved the Platform Commerce card leg to INVO's hosted checkout). Deprecations get a documentation-only notice in a minor release first — verify_sms_transfer/verify_sms_send are deprecated as of 3.1.0 and keep working until a future major removes them. It's at parity with the JS SDK (3.x) — same server surface, same webhook scheme, and the passkey-recovery relay — and the wire contract is the same live INVO API, backward-compatible within a major. Safe to depend on in production; pin a version and watch releases for updates.

Development

python -m venv .venv && . .venv/bin/activate      # (Windows: .venv\Scripts\activate)
pip install -e ".[dev]"
python -m pytest        # tests
python -m ruff check .  # lint
python -m mypy          # types (strict)

License

Proprietary — © Invo Tech Inc. See LICENSE.