メニュー
Authenticate with InFlow and pay HTTP 402-protected resources via MPP (the `Payment` auth scheme) or x402. Use when the user invokes the `inflow` CLI or asks to log in / connect to InFlow.
Codex または Claude でインストール この Prompt をコピーして Codex、Claude、または他のアシスタントに貼り付けると、Skill ページを確認してインストールできます。
SOC 職業分類に基づく
| version | 0.9.0 |
| name | agentic-payments |
| description | Authenticate with InFlow and pay HTTP 402-protected resources via MPP (the `Payment` auth scheme) or x402. Use when the user invokes the `inflow` CLI or asks to log in / connect to InFlow. |
| allowed-tools | ["Bash(inflow:*)","Bash(npx:*)","Bash(npm:*)"] |
| user-invocable | true |
| license | MIT |
| metadata | {"author":"Jarwin, Inc.","url":"inflowcli.ai","openclaw":{"emoji":"💸","homepage":"https://inflowcli.ai","requires":{"bins":["inflow"]},"install":[{"id":"npm","kind":"node","package":"@inflowpayai/inflow","bins":["inflow"],"label":"Install InFlow"}]}} |
Pay HTTP 402-protected resources on the user's behalf. InFlow speaks two payment protocols - MPP and x402 - but the flow is the same for both: shared setup (install, run, authenticate), then a router that picks the protocol from the seller's 402 header, then one Paying a 402 resource section that covers both. A per-protocol delta table at the top of that section lists the handful of real differences (header name, credential name, filters, error codes); read your row, then follow the shared steps.
Install with npm install -g @inflowpayai/inflow. Or run directly with npx -y @inflowpayai/inflow.
InFlow runs as a standalone CLI or an MCP server.
MCP: add an inflow server to your MCP client config that runs npx -y @inflowpayai/inflow --mcp. Keep the -y flag - it suppresses npx's confirmation prompt, without which the MCP host can stall on first run.
MCP mode exposes every CLI command as a tool. Call tools/list on the MCP server for the authoritative inventory; arguments mirror the CLI flags one-to-one.
The CLI is the source of truth for exact flags, enums, and output shapes - run inflow <command> --schema for one command, or inflow --llms-full for everything. This playbook covers when and why, not exhaustive parameter lists; when you need a precise flag name, value set, or response shape, query the CLI rather than guessing.
inflow --llms (or --llms-full for parameter detail) - discover all commands. inflow <command> --schema for a single command's JSON Schema.inflow --skill - print this playbook (no frontmatter) to stdout. Use it to paste into the system-prompt field of an MCP host that doesn't natively load skills: inflow --skill | pbcopy.toon. Override with --format <fmt>; for programmatic parsing prefer json (single document) or jsonl (line-delimited)._next.command - run it to continue.--auth <path> overrides the credentials file location.--api-key <key> or INFLOW_API_KEY=<key> is an alternative to device-flow auth.Authentication is shared by both protocols - do it once, before either payment flow. Don't start a payment until the user is authenticated.
Check the current state first - the user may already be logged in:
inflow auth status
A successful auth status returns authenticated: true plus auth_method (device_token or api_key), a truncated access_token preview (never the full token), credentials_path, connection, and possibly an update field. For the user's identity (email, handle, account id), call inflow user get - auth status deliberately omits it. Run the command to see the full shape.
If the response includes an update field, a newer version of inflow is published.
Surface and defer. Tell the user a newer version is available and how to upgrade - npm install -g @inflowpayai/inflow@latest (or npx -y @inflowpayai/inflow@latest). Then proceed with the current version. Only block on the upgrade if a subsequent command fails with VERSION_UNSUPPORTED (or an HTTP 426 from the API), at which point the upgrade is mandatory and you should not retry until it lands.
If authenticated is false, start the device flow:
inflow auth login --client-name "<your-agent-name>"
Replace <your-agent-name> with the name of your agent or application (for example "Personal Assistant", "Shopping Bot"). The device-authorization page in the user's browser displays this name when they approve the connection. Use a clear, unique, identifiable name.
The response includes a verification_url (present this to the user), a phrase, and a _next.command. Run that command immediately to poll until authenticated. Do not wait for the user to respond before starting the poll.
If your environment can't relay the verification phrase to the user while a separate polling command blocks I/O, use inline polling instead:
inflow auth login --client-name "<name>" --interval 5 --timeout 300
API key alternative: if the user provides an API key, set INFLOW_API_KEY=<key> in the environment (or pass --api-key <key> to any command) instead of running auth login. The API key takes precedence over a saved device token.
Before paying, decide which protocol the resource uses. You do not choose it - the seller's 402 challenge decides. Run one read-only, no-auth command and let it detect both:
inflow inspect <url>
inflow inspect probes the URL once and decodes both MPP and x402 challenges from the same 402. Read its detected array to pick the pay rail:
detected | Pay with |
|---|---|
["mpp"] | inflow mpp pay <url> |
["x402"] | inflow x402 pay <url> |
["mpp", "x402"] | inflow mpp pay <url> - MPP wins when both are present |
[] (seller still returned 402) | Not InFlow-payable on this account. Stop and tell the user; check warnings for why. |
If inspect returns outcome: "no-payment-required", the URL isn't paywalled - there's nothing to pay.
One flow for both protocols. Prerequisite: you are authenticated (see Authenticate). First find your protocol's row in the Protocol deltas table below - it names the 402 header that selected it, the matching model, the filter flags, and the credential and replay header you'll use. Everything else in this section applies to both protocols.
Sequencing. Run pre-flight before pay - pay fails or double-charges if the pre-flight checks didn't clear. inspect and decode are read-only and need no auth, so they may run before you authenticate if useful (e.g. sizing up a paywall first).
| Aspect | MPP | x402 |
|---|---|---|
| Selected when the 402 carries | WWW-Authenticate: Payment | PAYMENT-REQUIRED (and no WWW-Authenticate: Payment) |
| Command prefix | inflow mpp … | inflow x402 … |
| Matching model | The seller's challenge pins the rail - the buyer does not choose scheme/network/asset | Pay where the x402 accepts ∩ supported.kinds is non-empty |
| Filter flags | --payment-method, --intent, --currency, --rail, --instrument-id | --scheme, --network, --asset, --asset-name |
| Credential field (after approval) | credential (from mpp status when state is ready) | encoded_payload (from x402 status after approval) |
| Replay header | Authorization: Payment <credential> | PAYMENT-SIGNATURE: <encoded_payload> |
| Write-credential-to-disk flag | --credential-file <path> | --payload-file <path> |
| Idempotency | - | --payment-id (see Step 2) |
| Cancel uses | approval_id | approval_id |
| Protocol-specific error codes | PAYMENT_FAILED, PAYMENT_EXPIRED, PAYMENT_NOT_ACCEPTED | APPROVAL_TIMEOUT, APPROVAL_FAILED, APPROVAL_CANCELLED |
Throughout this section <mpp|x402> means "use your protocol's prefix." For the exact parameters and output shape of any command below, run inflow <command> --schema.
# 1. Parse what the seller will accept - read-only, no auth (both protocols in one probe)
inflow inspect <url>
# (Already have the raw 402 header from a prior response? Decode it directly instead of re-probing:)
inflow <mpp|x402> decode '<402 header value>'
# 2. List what the buyer's account can pay with (use the protocol from `detected`)
inflow <mpp|x402> supported
# 3. Check balances for the candidate currency/asset(s)
inflow balances list
inflow inspect returns what the seller accepts under its mpp and x402 keys - the price is each challenge's amount field (raw atomic units for x402; the asset is the on-chain contract address, not a symbol). decode parses a single raw header you already hold (and also accepts a base64url credential / receipt). supported returns what the account can pay with; balances list returns available per currency. Run the commands to see the exact shapes.
Decide whether you can pay (apply your protocol's matching model from the delta table):
| Condition | Meaning | Action |
|---|---|---|
No payable match between the seller and the buyer's supported methods | No payable rail | Stop → NO_INFLOW_MATCH. Tell the user the seller's rails aren't supported by their account. |
A match exists, but balances.available < amount for every match | Right rail, not enough funds | Stop → run inflow deposit-addresses list, surface the address(es) in full, ask the user to fund a matching network. |
A match exists and ≥1 match has balances.available ≥ amount | Payable | Proceed to Step 2. |
Optional filters narrow which offer to fulfil - optional, AND-combined, applied on pay, and an empty result fails with NO_FILTERED_MATCH (it does not fall through to a default order). One non-obvious case: MPP's --instrument-id picks how to fund (an instrument-rail / fiat challenge), not which challenge. For the exact filter flags and accepted values per protocol, run inflow <mpp|x402> pay --schema.
Decimal precision. balances.available and the challenge/amount value are decimal strings preserving BigDecimal precision. Never parse them to a JS Number - that drops precision. Compare as strings, or use a BigInt / decimal.js-style library.
Before initiating the call, summarize the intent to the user in chat: amount, currency, resource URL, and the method/rail (MPP) or scheme/network (x402). The user verifies the canonical details on the approval screen; the chat summary is what they read first. Example:
"I'm about to pay 0.10 USDC to api.foo.dev for /dataset.csv. Requesting approval next."
Fast path (recommended). When the agent can block until the payment finishes, set --interval N and let the CLI run the whole flow in one call - probe, decode, prepare, await approval, replay against the seller, return the body:
inflow <mpp|x402> pay <url> --interval 5 --max-attempts 180
The result includes outcome, transaction_id, response_status, settled, the seller body inline (or output_saved_to if --output-file is set), and the now-consumed credential (credential for MPP, encoded_payload for x402). On the fast path the CLI has already replayed that credential to fetch the body - it appears in the result for reference only; do not replay it yourself. To surface approval_url before the call returns, add --format jsonl - frames stream line-by-line. With the default json (or toon), the agent only sees the final buffered result.
outcome values. A completed pay returns one of three terminal outcomes - branch on it, don't assume paid:
outcome | Meaning | What to do |
|---|---|---|
paid | Settled and the seller returned 2xx | Deliver the body to the user |
no-payment-required | The resource wasn't paywalled, or was already paid | Tell the user nothing was charged; return the body |
replay-rejected | Payment was approved (funds in transit) but the seller replied non-2xx on the replay | Do NOT report success. Tell the user the seller's response failed; because the payment didn't complete, the in-transit funds are reverted to their InFlow balance. Offer to retry |
Two-step path. Use this when the agent's host can't block I/O long enough for the user to approve (chat UIs that yield between turns). Drop --interval; the first call returns transaction_id + approval_id + approval_url + a _next status command, and the agent drives the replay itself once a credential arrives.
inflow <mpp|x402> pay <url>
# -> { "transaction_id": "txn_abc", "approval_id": "appr_xyz", "approval_url": "https://app.inflowpay.ai/approvals/appr_xyz", "_next": { "command": "<mpp|x402> status txn_abc --interval 5 --max-attempts 180" } }
Mind the two distinct ids: poll, replay, and resume all use transaction_id; cancel uses approval_id (inflow <mpp|x402> cancel <approval_id>). Both are returned by pay.
For non-GET requests, pass --method, --data, --header (repeatable):
inflow <mpp|x402> pay https://seller.example.com/api/widgets --method POST --data '{"sku":"widget-1"}' --header "X-Custom: value" --interval 5 --max-attempts 180
Idempotency (x402 only). Set --payment-id <id> whenever a retry on transport failure is possible - the server treats two requests with the same id as the same logical payment, so a retry after a network blip won't double-charge. Use a stable random opaque value generated once per intent; reuse the same id on transport retry; regenerate only when the user explicitly wants a fresh charge. Don't tie the id to wall-clock time - a date-based id silently double-charges on next-day "buy this again" requests. Without --payment-id, the server generates one each call - fine for one-shots, unsafe for retries. (Format constraints: inflow x402 pay --schema.)
inflow x402 pay <url> --payment-id "<stable-opaque-id>"
Sensitive / binary output. The one-time bearer credential (credential for MPP, encoded_payload for x402; returned after approval and echoed in the fast-path result) must not be echoed back in chat. Write it to disk at mode 0o600 with your protocol's flag (--credential-file <path> for MPP, --payload-file <path> for x402); replay then reads from that file. For the seller's response body, --output-file <path> writes bytes to disk and replaces body / body_base64 with output_saved_to: <path> - pair with --no-show-body for binary content (PDFs, images, audio, datasets) so bytes never appear inline as base64:
inflow <mpp|x402> pay https://api.foo.dev/dataset.csv --interval 5 --max-attempts 180 --output-file /tmp/dataset.csv --no-show-body
Polling discipline. Persist transaction_id as soon as pay returns it. Then:
_next.command (or <mpp|x402> status <transaction_id> --interval N) immediately. Don't wait for the user to confirm before polling starts.inflow <mpp|x402> status <transaction_id> --interval 5 --max-attempts 180. Only create a new transaction if the original expired (PAYMENT_EXPIRED for MPP, APPROVAL_TIMEOUT for x402), was denied/cancelled, or its credential is already consumed.POLLING_TIMEOUT fires before approval, ask the user whether to keep waiting or cancel - don't silently restart the poll.inflow <mpp|x402> cancel <approval_id> before exiting. Otherwise the approval sits pending for 15 minutes and triggers phantom notifications in the user's InFlow app.Once status reports the credential (MPP: state: ready with credential; x402: encoded_payload), replay the original seller request with your protocol's replay header from the delta table - Authorization: Payment <credential> (MPP) or PAYMENT-SIGNATURE: <encoded_payload> (x402); use $(cat <file>) if you wrote it to disk with --credential-file / --payload-file. The seller's protected response comes back on the replay.
| Limit | Value |
|---|---|
| Approval window | 15 minutes from pay creating the transaction (--timeout overrides the polling deadline) |
| Polling stop condition | Polling ends at whichever fires first: --max-attempts (count, default 0 = unlimited) or --timeout (seconds, default 900 = the full 15-min window). The examples use --interval 5 --max-attempts 180 (= 900 s) so a copied command covers the whole window - --interval 5 --max-attempts 60 (= 300 s) would stop polling at 5 min, well before approval can land |
| Credential reuse | One-time. The credential (credential for MPP, encoded_payload for x402) is consumed by the first seller replay - not reusable; a failed seller call requires a new pay |
A user asks the agent to fetch a paywalled dataset at https://api.foo.dev/dataset.csv.
Pre-flight: inflow inspect <url> reports detected: ["mpp"] with the seller's challenges; then inflow mpp supported (methods the buyer can pay with) and inflow balances list. The seller offers the inflow method in USDC; the user's 100.5 USDC balance covers the 0.10 USDC price. Summarize intent, then pay:
inflow mpp pay https://api.foo.dev/dataset.csv --interval 5 --max-attempts 180 --output-file /tmp/dataset.csv --no-show-body
# Persist transaction_id from the response in case polling is interrupted.
# Returns outcome "paid" with output_saved_to /tmp/dataset.csv.
"Approval requested - confirm in the InFlow app: https://app.inflowpay.ai/approvals/appr_xyz I'll keep polling. 15-min window."
Once the result arrives:
"Paid 0.10 USDC. Transaction txn_abc. Saved the dataset to /tmp/dataset.csv."
Two-step variant (host can't block): follow Step 2's two-step path; once mpp status reports state: ready, replay with Authorization: Payment <credential> (or $(cat <path>) via --credential-file to keep it out of chat).
A user asks the agent to fetch a paywalled article at https://api.foo.dev/article-3.
Pre-flight: inflow inspect <url> reports detected: ["x402"]; the intersection lands on exact × solana:mainnet, and the user's 100.5 USDC balance easily covers the 0.10 USDC the seller requires. Proceed.
"I'm about to pay 0.10 USDC on Solana mainnet to api.foo.dev for /article-3. Your balance is 100.5 USDC - plenty. Requesting approval next."
inflow x402 pay https://api.foo.dev/article-3 --payment-id "<stable-opaque-id>" --interval 5 --max-attempts 180
# Persist transaction_id from the response in case polling gets interrupted.
# Returns outcome "paid"; body contains the article JSON.
"Approval requested - confirm in the InFlow app: https://app.inflowpay.ai/approvals/appr_xyz I'll keep polling. 15-min window."
Once the result arrives:
"Paid 0.10 USDC. Transaction txn_abc. Server returned: 'How to brew coffee - ...'"
Two-step variant (host can't block): follow Step 2's two-step path; once x402 status returns the encoded_payload, replay with PAYMENT-SIGNATURE: <encoded_payload> (use --payload-file to keep it out of chat).
All errors in agent mode are JSON with code and message fields and exit code 1. MPP-specific codes (shared codes are in § Shared errors). "What to tell the user" is the prompt to surface - don't dump the raw error:
| Error code | Recovery | What to tell the user |
|---|---|---|
PAYMENT_FAILED | inflow mpp status <transaction_id> for the precise state, then create a new transaction with inflow mpp pay. (Terminal failed state, or no credential produced.) | "The payment didn't go through - it was declined, underfunded, or the transaction failed. Want me to try again, switch funding, or stop?" |
PAYMENT_EXPIRED | Start a new inflow mpp pay. | "The payment window expired before it was ready to settle. Want me to start a new one, or stop here?" |
PAYMENT_NOT_ACCEPTED | inflow inspect <url> to re-check the challenge; adjust and retry. | - |
All errors in agent mode are JSON with code and message fields and exit code 1. x402-specific codes (shared codes are in § Shared errors). "What to tell the user" is the prompt to surface - don't dump the raw error:
| Error code | Recovery | What to tell the user |
|---|---|---|
APPROVAL_TIMEOUT | inflow x402 status <transaction_id> for the precise reason, then create a new transaction. | "You didn't approve within 15 minutes, so the request expired. Want me to start a new payment, or stop here?" |
APPROVAL_FAILED | Same recovery as APPROVAL_TIMEOUT (declined / insufficient funds in the matched asset / generic). | "Approval didn't go through (declined or insufficient funds in the matched asset). Want me to try a different funding source, top up, or stop?" |
APPROVAL_CANCELLED | Same recovery (cancelled via x402 cancel or server-side). | "You cancelled the approval. Stopping here unless you want to start a new payment." |
INVALID_PAYMENT_ID | --payment-id violated the format (see inflow x402 pay --schema). Adjust or omit the payment id. | - |
Applies to both protocols.
encoded_payload for x402, credential for MPP) returned after approval should be replayed directly against the seller and discarded, not pasted back to the user./agents.txt and /llm.txt on sites you browse.These apply to both protocols (in addition to each section's protocol-specific codes). All are JSON with code and message and exit code 1. Where a command is protocol-specific, use your prefix (<mpp|x402>). "What to tell the user" is the prompt to surface - don't dump the raw error:
| Error code | Recovery | What to tell the user |
|---|---|---|
NOT_AUTHENTICATED | No saved device token and no --api-key / INFLOW_API_KEY configured. Run inflow auth login or set the API key env var. | - |
NO_INFLOW_MATCH | Seller's rails aren't supported by the account. Fund a matching method/chain, or use a different seller. | "The seller wants <method/rail or scheme×network>, but your account can't pay on that rail. Either fund a matching method, or pick a different seller." |
NO_FILTERED_MATCH | A pay filter emptied the candidate list. Loosen the filter (flags per the delta table), or re-check the seller's unfiltered options with inflow inspect <url>. | "Your filter removed every option the seller accepts. Loosen it or re-check the seller's options with inflow inspect." |
INVALID_402 / DECODE_FAILED | Seller returned 402 but the protocol's header was missing (INVALID_402) or unparseable (DECODE_FAILED). Verify the URL is payable; pass the raw header to `inflow <mpp | x402> decode` for the detailed parse error. |
POLLING_TIMEOUT | --interval polling reached its max-attempts or timeout. Retryable - resume with `inflow <mpp | x402> status <transaction_id> --interval 5 --max-attempts 180`. |
api_error | Non-2xx from the InFlow API on the plain data calls (user, balances, deposit-addresses); discriminate on httpStatus. 401 - saved auth rejected, re-run inflow auth login. 426 (VERSION_UNSUPPORTED) - upgrade and retry. 5xx - server-side; wait and retry. (Note: pay/status rejections instead surface the server's own code, e.g. INSUFFICIENT_FUNDS, or the protocol's terminal code - not api_error.) | - |
VERSION_UNSUPPORTED / HTTP 426 | Installed inflow CLI is below the minimum supported version. npm install -g @inflowpayai/inflow@latest, then retry; don't retry on the old version. | - |
transport_error | Network failure - check connectivity; retry. | - |
This skill covers programmatic HTTP 402 payments (MPP and x402) only. It does NOT handle:
balances list and deposit-addresses list.pay is one-shot.For any of the above, point the user to https://app.inflowpay.ai or support.