> ## Documentation Index > Fetch the complete documentation index at: https://docs.flow.engineer/llms.txt > Use this file to discover all available pages before exploring further. # Show the typing indicator > | Channel | Works | Limits | |---|---|---| | Telegram | Yes | | | WhatsApp | Yes | within 24 hours of the contact's last message, after they wrote at least once | | iMessage | Yes | within 5 minutes of the contact's last message; not in group conversations; `imessage.duration_seconds` 1 to 60 | | SMS (US, pilot) | No | | On SMS, send instead: nothing (SMS has no typing indicator); send your reply when it is ready. Shows the typing indicator in a conversation. The body is optional: no body or `{}` shows it. No channel lets it be cleared on request: it clears itself when your next message goes, or after a few seconds, so call again every few seconds while your agent works. `Idempotency-Key` works as on every `POST`: a repeat with the same key within 24 hours answers the first call's result and does not show the indicator again, so give each call that should show it its own key (or none). SMS answers `422 unsupported_content` with `param` `kind`, and an iMessage group conversation with `param` `kind` and `channel_code` `group_unsupported`: nothing is sent, so leave typing out there. The call goes to the channel at once, not behind the conversation's queued messages, so the channel's answer is the call's answer. It fails with `409 outside_window` when the channel allows typing only inside a window (iMessage: within 5 minutes of the contact's last message; WhatsApp: within 24 hours of it, and only after the contact wrote at least once), and with `502 channel_error` when the channel failed or timed out. Both are safe to ignore: typing is a courtesy, so never hold back a reply because of it. ## OpenAPI ````yaml unified-messaging-api/openapi.yaml POST /v0/conversations/{conversation_id}/typing openapi: 3.1.0 info: title: Flow Messaging API version: '2026-11-01' summary: Two-way messaging for AI agents on WhatsApp, Telegram and iMessage. description: > Flow Messaging lets an AI agent hold conversations with people on **WhatsApp**, **Telegram** and **iMessage** through one HTTP API, and on **SMS and MMS in the US** (in pilot: dedicated numbers arranged with the Flow team). Flow hosts the senders (bots, numbers and lines), receives every inbound message, and passes every send through one gate that enforces each channel's rules. You never handle vendor keys or channel webhooks. Channels today: Telegram is live. iMessage is live on lines the Flow team connects to your app: it replies to people who write first, in one-to-one conversations and in group chats people add the line to; writing first is available on request on dedicated lines (`Sender.imessage.can_initiate`). WhatsApp numbers are connected by the Flow team too, either a number Flow provides or one in your business's own WhatsApp Business Account, and the WhatsApp-only parts (templates, the 24-hour window) apply on them; the shared WhatsApp sandbox is not open yet. This API is in **beta**: the shape may still change before it is declared stable. Changes are released as new dated versions (see Versioning). ## Concepts - An **app** is one agent integration. It owns API keys, webhook endpoints and settings, and has a test mode and a live mode. - A **sender** is what your agent talks *from*: a Telegram bot, a WhatsApp number, an iMessage line or a US SMS number. Shared sandbox senders serve many apps; dedicated senders are yours alone. - A **contact** is a person on one channel. There is no cross-channel identity: a contact on Telegram and a contact on WhatsApp are different contacts, even if they are the same person. Never assume a contact has a phone number. - A **conversation** is one sender talking with one contact (`type` `direct`), or, on iMessage, with a group chat the line was added to (`type` `group`, see Groups). You reply into a conversation; you never pick a channel per message. - A **message** is one **kind**: `basic` (text, or media with a caption), `voice`, `buttons`, `list`, `carousel`, `template`, `contact_card`, `location`, `effect`, `reaction`, and more (see `MessageKind`). Each kind you can send has its own endpoint, named after it (`POST /v0/conversations/{conversation_id}/basic`, `/buttons`, ...), whose page opens with the channels it works on. The same kind objects describe what you receive (`Message.content`, with `kind`). - An **event** is one entry in your app's ordered log. Everything you receive (inbound messages, delivery statuses, reactions, sender changes) is an event, delivered by webhook, by the live stream (`GET /v0/stream`) and by polling (`GET /v0/events`). ## Authentication Every request carries an API key as a bearer token: ``` Authorization: Bearer fk_live_... ``` Keys belong to one app and one mode. `fk_test_` keys see only sandbox senders and test data, and nothing they send reaches a real contact unless that contact joined the sandbox. `fk_live_` keys reach real contacts through your dedicated senders. A key is shown once when it is created and stored only as a hash. Keys are never accepted in the query string. The live stream (`GET /v0/stream`) also takes the key as a WebSocket subprotocol, for browsers, which cannot set headers on a WebSocket (see that endpoint). ### No key yet? Get one in one call Anyone, including an AI coding agent, can get a test key without signing up: ``` curl -X POST https://api.flow.engineer/v0/sandbox/keys ``` The answer (`SandboxKey`) holds a `fk_test_` key for a new app of its own, shown once, and the sandbox senders with the links and join message a person uses to join it. Save the key (for example as `FLOW_MESSAGING_KEY`) and the `claim_token`. This key needs no account and has a **sandbox allowance**: 1 contact and 50 messages sent in total on the Telegram sandbox (and WhatsApp's when it opens), and it **expires after 7 days**. Only messages your agent sends count; inbound messages are free. iMessage is not part of the allowance. To keep the app, a person signs in with a Flow account (flow.engineer) through the device flow: the agent or the CLI (`npx @flow-engineer/messaging login`) calls `POST /v0/device/authorizations` with the `claim_token`, shows the person the `verification_uri` and `user_code` (they type the code on that page), and polls `POST /v0/device/token` until it returns a key. The app is claimed: its data and keys are kept, its keys no longer expire, and it moves under the person's signed-in allowance of 3 contacts and 100 messages each, **one allowance per person, shared by every app they own or claim** (a person may claim up to 10 apps). `GET /v0/app` shows what is left of the allowance (`allowance`). When it runs out, sends answer `403 permission` with `channel_code` `sandbox_allowance_used`. The apps of your own account get the free trial instead, started at your first sign-in: 25 people per channel on Flow's shared Telegram, WhatsApp and iMessage senders, over all the account's apps (their `allowance` is absent). ## Versioning The API is versioned by date. Send `Flow-Version: 2026-11-01` to choose a version; without the header, the version pinned to your app when it was created is used. Responses echo the version that served them in `Flow-Version`. Every path starts with `/v0`: this is API V0, the shape before the stable release. `/v1` is kept for that release, version 1.0, after which the shape is frozen. ## Idempotency Every `POST` that takes an API key accepts an `Idempotency-Key` header (any unique string up to 255 characters; a UUID or ULID is a good choice); the onboarding calls `POST /v0/sandbox/keys`, `POST /v0/device/authorizations` and `POST /v0/device/token` ignore it. Keys are kept for 24 hours per app and mode. Repeating a request with the same key returns the first answer, with the header `Idempotent-Replayed: true`, and does not act twice. Otherwise a key already used answers `409 idempotency_conflict`, and its `channel_code` says which case it is: - `body_mismatch`: the key was used with a different method, path or body. Use a new key for a new request; do not retry. - `in_progress`: the first request with this key is still running. Wait `retry_after` seconds and repeat the identical request with the same key to get its answer. - `secret_not_kept`: the first request went through, but its answer carried a secret shown only once (creating a webhook endpoint, rotating its secret, connecting a WhatsApp number), so that answer was not kept. Answers that are worth retrying (`429`, `500`, `502`, `503`, `504`) are not kept, so a retry with the same key runs again. ## Errors Errors share one shape: ```json {"error": { "type": "unsupported_content", "channel": "telegram", "param": "kind", "message": "Carousels work on WhatsApp only.", "hint": "On Telegram, send a basic message, then one buttons message with an image header per card.", "doc_url": "https://api.flow.engineer/docs/errors/unsupported_content", "conversation": "conv_01JB..."}} ``` `error.type` is a closed list (see the `ErrorType` schema); switch on it, not on the message text. Every error carries `hint`, one sentence saying what to change for this case, and `doc_url`, the page for its type (`https://api.flow.engineer/docs/errors/`: what it means, why it happens, how to fix it, with code). Errors that can be retried carry `retry_after` in seconds. Coding agents can act on `hint` directly; the MCP tool `explain_error` returns the page for a type. ## Pagination Lists are paged with cursors. Pass `limit` (1 to 100, default 20) and either `after` or `before`, set to the ID of an item you already have. IDs are prefixed ULIDs, so they sort by creation time. The event log (`GET /v0/events`) is oldest first, so `after` moves forward in time; every other list is newest first, so `after` moves back in time. A list answer says `has_more` when more items lie in that direction. ## Rate limits Requests are limited per key. Every answer carries `RateLimit-Limit`, `RateLimit-Remaining` and `RateLimit-Reset`. Sends are also paced per sender by the send gate (see below): each sender has a sending rate it may not exceed. Going over either limit answers `429` with type `rate_limited`, a `Retry-After` header and `error.retry_after` in seconds; wait that long and retry with the same `Idempotency-Key`. The send gate also answers `429` with type `new_contact_limit` when a sender has used its budget for starting conversations. ## The send gate Every send (HTTP, the stream, or a reply in a webhook response) passes the same gate, which also refuses what the channel does not take (see "Exact or 422"): - **WhatsApp**: free-form messages only while the contact's last message is less than 24 hours old (the window); otherwise only a `template` (`409 outside_window`). - **iMessage**: a line may only reply unless it can start conversations (`Sender.imessage.can_initiate`, available on request on dedicated lines); a new conversation from a reply-only line is refused (`403 permission`). A line that can start conversations spends the sender's new-contact budget, with a warm-up curve, and a contact who has never messaged the line (or opted in) gets a `message.failed` event with `outside_window`. Group conversations always begin with the group writing; a start never makes one. - **Telegram**: the contact must have started the bot; Telegram enforces this. - **SMS (US, in pilot)**: a conversation starts only with a contact who texted one of your numbers before, or whose consent you record with the start (`consent` on `POST /v0/messages`); otherwise `403 permission` with `channel_code` `consent_required`. A contact who texts STOP (or STOPALL, UNSUBSCRIBE, CANCEL, END, QUIT, REVOKE, OPT OUT, OPTOUT) is opted out of every SMS sender of your app at once: sends to them answer `403 permission` with `channel_code` `contact_opted_out`, and messages already queued fail the same way, until they text START (or UNSTOP, YES) to a number they opted out through (carriers and providers block per number, so START to another of your numbers changes nothing). HELP (or INFO) gets the sender's help reply. Flow answers these keywords itself, and you still receive the message as `message.received`. The contact's `opted_out_at` shows the state. Messages to a contact who has not texted in the last hour respect the sender's quiet hours (`Sender.sms.quiet_hours`, by default 21:00 to 08:00 in the contact's time zone, worked out from the area code): inside them a send answers `409 outside_window` with `channel_code` `quiet_hours` and `retry_after` set to when they end; a message that waited in the queue into the quiet hours is held until they end, not sent at night. - **HIPAA accounts** (`Account.hipaa`): only channels under a business associate agreement carry messages (SMS). Sends, typing and read receipts on Telegram, WhatsApp and iMessage, and connecting senders there, answer `403 permission` with `channel_code` `hipaa_channel_not_covered`; messages already queued there fail the same way. Contacts get a fixed refusal from the sandbox senders instead of reaching the app, and nothing they send there is stored or delivered. Simulated conversations (`POST /v0/test/conversations`) follow the same rule: only SMS can be simulated. - **Sandbox senders**: only contacts who joined through your app (by sending its join message, for example `Hi Sandbox app (ref brave-otter-40718263)`, or by opening its Telegram link). Starting a conversation spends budget per sender (new contacts per day and per hour, with a warm-up curve for new lines) and per plan. When abuse signals trip (the same text to many new contacts, many starts with no reply, blocks, a drop in WhatsApp quality) the sender is throttled and you receive `sender.status_changed`: until the sender's `throttled_until`, starts fail with `429 sender_throttled`, while replies into existing conversations still go. The sender recovers by itself, with another `sender.status_changed`. ## Groups (iMessage) When someone adds one of your dedicated iMessage lines to a group chat (Flow's shared iMessage line ignores group chats), the group is one conversation with `type` `group`. It has no `contact`; `group` lists the participants (`participants`, each a Flow contact on the channel, made the first time Flow sees them) and the group's name when it has one. Every inbound message carries `from`, the contact who wrote it. Participants who join or leave change the list without an event: read the conversation (`GET /v0/conversations/{conversation_id}`) for the current list. Reply with the normal sends (`POST /v0/conversations/{conversation_id}/basic` and the other kinds' endpoints). Basic messages (with `imessage.styles`, `imessage.subject` and `link_preview`), voice notes, effects, `reply_to`, edits and unsends work in groups. Groups take no reactions, typing indicator or read receipts from your line: each answers `422 unsupported_content` with `param` `kind` and `channel_code` `group_unsupported`, and nothing is sent. Dedicated iMessage lines receive and reply in groups; Flow's shared iMessage line ignores group chats. A group cannot be started, renamed, or have people added or removed through the API. ## Exact or 422 What you send is what the contact gets. Each endpoint's page opens with the channels it works on and their limits; a send of a kind, a field, a channel block or a value the conversation's channel does not take fails with `422 unsupported_content` and sends nothing. `param` names what was refused (`kind`, a field such as `reply_to` or `telegram.is_big`, or a value's field such as `emoji`), `channel` the conversation's channel, and `hint` what to send instead over plain HTTP (for example, for a carousel on Telegram, a basic message, then one buttons message with an image header per card). Flow never converts a message into another kind, never drops a field, and never answers `2xx` for something it did not send. Fields that only one channel has sit in a block named after it (`telegram`, `imessage`), typed per endpoint; an unknown key in a block, or an unknown field, is refused with `400 invalid_request` naming it, and so is a read-only field Flow fills on received messages (such as `media.mime_type` or a voice note's `transcript`). `text` is sent exactly as written on every channel: Flow does not read markdown, strip marks or warn. Style it in the channel's own terms: `telegram.parse_mode`, `imessage.styles`; WhatsApp's app reads its own `*bold*` and `_italic_`. With `telegram.parse_mode`, Telegram's length limits count the text after its tags or marks are parsed, as Telegram counts them, so a styled text may be longer than the limit before parsing. ## Order and duplicates Sends to one conversation go out in the order Flow accepted them, one at a time, across every endpoint, webhook replies and the stream. A message sent as several (a text and one message per card) arrives in the order you sent the calls. Give each call its own `Idempotency-Key` (for example `evt_123:card:2` for the second card answering event `evt_123`): a retry with the same key never sends twice (see Idempotency). A `422` answer is kept under its key like any other, so a message you send instead of a refused one (a `basic` after a `buttons` came back `422`) is a new request with a new key: add what it is to the refused one's key, such as `evt_123:card:2:basic`. ## Webhooks Register endpoints with `POST /v0/webhook_endpoints`. Each event is `POST`ed as JSON (the `event` webhook below), in order per conversation, and signed: ``` Flow-Signature: t=1791763200,v1=5257a869e7ecebeda32affa62cdca3fa51cad7e77a0e56ff536d0ce8e108d8bd ``` `v1` is the lowercase hex HMAC-SHA256, keyed with the endpoint's signing secret (`whsec_...`), of the string `t.{t}.{body}`: the letter `t`, a dot, the timestamp from the header, a dot, and the raw request body exactly as received. Compare in constant time and reject timestamps more than 5 minutes from your clock. While a secret is being rotated (`POST /v0/webhook_endpoints/{webhook_endpoint_id}/rotate_secret`), the header carries one `v1` per active secret; accept the request if any of them matches. Answer `2xx` within 10 seconds. To reply at once, answer a `message.received` event with `200` and a body `{"reply": }` or `{"reply": [, ...]}` (up to 11 pieces, each a kind object with `kind`, such as `{"kind": "buttons", ...}`); the reply is sent into the event's conversation through the send gate, in order, which saves a round trip. A string is a basic message: `{"reply": "Hi"}` sends `{"kind": "basic", "text": "Hi"}`. Any other `2xx` answer (an empty body, plain text such as `OK`, or JSON without `reply`) sends nothing and is not an error. A JSON object whose `reply` is not valid `WebhookReply` content, or a body that starts with `{` but is not valid JSON, sends nothing at all and is recorded on the delivery as an `invalid_request` error (the MCP tool `get_webhook_deliveries` shows it); the delivery still counts as delivered and is not retried. Slow agents answer `200 {}` at once and send later with the kinds' endpoints (`POST /v0/conversations/{conversation_id}/basic`, ...). Failed deliveries are retried with backoff for 3 days; within a conversation, later events wait behind a failing one. After that the event stays in the log, marked failed, and can be read again with `GET /v0/events`. Inbound messages keep the channel's order within a conversation. iMessage: inbound order is best-effort (arrival order); messages sent within about a second of each other may arrive out of order. ## Retention and deleting contacts Message content and files are kept for your plan's retention period (30 days by default). After it, a message's content is removed and the message is no longer returned, files are deleted, and events older than it leave the log. `DELETE /v0/contacts/{contact_id}` erases one contact at once, for example when the person asks you to delete their data or when you delete them in your own product: their name and address, the content of every message in their conversations, the files in them, their events and pending webhook deliveries, and messages still queued to them. It cannot be undone. What stays is only IDs, timestamps and the counts that bill and pace your senders. If the person writes again, or you send to their address, they are a new contact with a new ID. ## Testing with simulated contacts A test key can play the end user's side without a phone or a real channel. `POST /v0/test/conversations` creates a conversation on `telegram`, `whatsapp`, `imessage` or `sms` with a **simulated contact**; `POST /v0/test/conversations/{conversation_id}/inbound` writes as that contact (text, or media uploaded with `POST /v0/files`). Each turn reaches your agent exactly like a real message: a `message.received` event by webhook, on the stream and in the event log, in order. Your agent replies with its normal sends, which pass the real send gate (the channel's kinds and fields, text limits, edit targets) and end `sent` on your app's simulated sender, without touching a network. On Telegram, as with a real bot, nothing reaches a contact who has not written yet: a send before the contact's first turn ends `failed` (`channel_error`, `channel_code` `403`). `GET /v0/test/conversations/{conversation_id}/transcript` shows both sides, oldest first: what was sent is exactly what the contact sees. Simulated traffic is free: it does not count against the sandbox allowance, since nothing leaves Flow and no real person is reached. Two limits bound it instead: an app has at most 500 simulated conversations (a new one past that answers `403 permission` with `channel_code` `simulated_conversation_limit`), and a key makes at most 120 requests a minute to the `/v0/test/` endpoints (`429 rate_limited`), counted on top of its plan's request rate limit. The gate also skips the rules that protect a real number's standing with its channel (pacing, the new-contact budget and warm-up, abuse throttling), so a test suite can run cases in parallel. Simulated conversations, contacts and senders carry `simulated: true` (senders `kind: simulated`), exist only in test mode, are never shown to live keys, and follow your plan's retention like other test data. The SMS rules that concern the contact hold on a simulated SMS conversation as on a real number. A start or reply before the contact's first turn needs their recorded consent (`consent` on `POST /v0/messages`; otherwise `403 permission` with `channel_code` `consent_required`). A turn with the text STOP (or a variant) opts the simulated contact out (its `opted_out_at`, `opted_out_via` `keyword`; sends answer `403 permission` with `channel_code` `contact_opted_out`), and START to the simulated sender lifts it; the turn is still delivered as `message.received`, and the transcript shows no automatic keyword reply. A simulated SMS sender keeps no quiet hours, so a test's outcome never depends on the time of day, and it sends media as MMS. A HIPAA account (`Account.hipaa`) can simulate only SMS: creating a simulated `telegram`, `whatsapp` or `imessage` conversation, or posting a turn to one made before the flag was set, answers `403 permission` with `channel_code` `hipaa_channel_not_covered`. A simulated WhatsApp conversation keeps WhatsApp's rules. The contact's turns go through the same inbound path as a real WhatsApp message, so each turn opens the 24-hour window, and the send gate applies WhatsApp's rules: outside the window (also before the contact's first turn) only a `template` can be sent (`409 outside_window`); buttons take up to 3 reply buttons (a `list` up to 10 rows); reactions and `reply_to` work; edits and unsends answer `422 unsupported_content`. ## MCP server `https://api.flow.engineer/mcp` is a Model Context Protocol server (Streamable HTTP transport, protocol version 2025-06-18 or later) for coding agents such as Claude Code, Codex and Cursor. It is not a REST endpoint: clients `POST` JSON-RPC messages to it (and `GET` answers `405`, since the server keeps no sessions). Authenticate with the same API key as the REST API: ``` claude mcp add --transport http flow https://api.flow.engineer/mcp --header "Authorization: Bearer $FLOW_MESSAGING_KEY" ``` Without a key, `/mcp` answers `401` and its `hint` gives the one call that gets a test key (`curl -X POST https://api.flow.engineer/v0/sandbox/keys`). The key's mode chooses the tools. A `fk_test_` key gets build-time tools to integrate and verify an app end to end: `whoami`, `sandbox_join`, `send_test_message`, `simulate_inbound`, `get_test_transcript`, `wait_for_event`, `list_events`, `get_webhook_deliveries`, `replay_event` and `explain_error`. A `fk_live_` key gets production tools: `send_message`, `reply`, `react`, `edit_message`, `unsend_message`, `typing`, `list_conversations` and `get_conversation_messages`, plus `whoami` and `explain_error`. The tools that send take a `message` that is a kind object, as in a webhook reply (a string is a basic message's text). Every send passes the same send gate as the REST API, and a refusal comes back as the tool's error with the same `type`, `param`, `hint` and `doc_url`. contact: name: Flow Engineer url: https://docs.flow.engineer license: name: Apache-2.0 url: https://www.apache.org/licenses/LICENSE-2.0 servers: - url: https://api.flow.engineer description: Production. Test and live mode are chosen by the API key. security: - bearerAuth: [] tags: - name: Messages description: >- Send messages into conversations, one endpoint per kind; start conversations; edit and unsend messages. - name: Conversations description: >- One sender talking with one contact. Read history, show typing, mark as read. - name: Events description: >- The ordered log of everything that happened in your app. Catch up, replay and debug. - name: Stream description: Live events over a WebSocket, resumable from any event. - name: Files description: >- Media in and out. Flow fetches channel media for you and serves it through short-lived URLs. - name: Senders description: >- The Telegram bots, WhatsApp numbers, iMessage lines and SMS numbers your app sends from. - name: Templates description: WhatsApp message templates, mirrored from Meta with their approval status. - name: Webhook endpoints description: Where Flow delivers your app's events. - name: Contacts description: People your app talks with, one per channel. - name: App description: The app and key making the request. - name: Testing description: > Test keys only. Simulated contacts: create a conversation with an end user who exists only inside Flow, write as them, and read what your agent sent back, with no phone and no real channel. - name: Onboarding description: > Get a test key without an account (`POST /v0/sandbox/keys`), and sign in with a Flow account (flow.engineer) through the device flow to claim the app and lift its sandbox allowance. These endpoints take no API key. paths: /v0/conversations/{conversation_id}/typing: parameters: - $ref: '#/components/parameters/ConversationId' - $ref: '#/components/parameters/FlowVersion' post: tags: - Conversations summary: Show the typing indicator description: > | Channel | Works | Limits | |---|---|---| | Telegram | Yes | | | WhatsApp | Yes | within 24 hours of the contact's last message, after they wrote at least once | | iMessage | Yes | within 5 minutes of the contact's last message; not in group conversations; `imessage.duration_seconds` 1 to 60 | | SMS (US, pilot) | No | | On SMS, send instead: nothing (SMS has no typing indicator); send your reply when it is ready. Shows the typing indicator in a conversation. The body is optional: no body or `{}` shows it. No channel lets it be cleared on request: it clears itself when your next message goes, or after a few seconds, so call again every few seconds while your agent works. `Idempotency-Key` works as on every `POST`: a repeat with the same key within 24 hours answers the first call's result and does not show the indicator again, so give each call that should show it its own key (or none). SMS answers `422 unsupported_content` with `param` `kind`, and an iMessage group conversation with `param` `kind` and `channel_code` `group_unsupported`: nothing is sent, so leave typing out there. The call goes to the channel at once, not behind the conversation's queued messages, so the channel's answer is the call's answer. It fails with `409 outside_window` when the channel allows typing only inside a window (iMessage: within 5 minutes of the contact's last message; WhatsApp: within 24 hours of it, and only after the contact wrote at least once), and with `502 channel_error` when the channel failed or timed out. Both are safe to ignore: typing is a courtesy, so never hold back a reply because of it. operationId: setTyping parameters: - $ref: '#/components/parameters/IdempotencyKey' requestBody: required: false content: application/json: schema: $ref: '#/components/schemas/TypingMessage' examples: show: summary: Show the indicator value: {} imessage_longer: summary: Show it for 30 seconds on iMessage value: imessage: duration_seconds: 30 responses: '202': description: The channel showed the indicator. content: application/json: schema: $ref: '#/components/schemas/ConversationAction' '400': $ref: '#/components/responses/InvalidRequest' '401': $ref: '#/components/responses/Unauthenticated' '403': $ref: '#/components/responses/PermissionDenied' '404': $ref: '#/components/responses/NotFound' '409': $ref: '#/components/responses/Conflict' '413': $ref: '#/components/responses/PayloadTooLarge' '422': $ref: '#/components/responses/UnsupportedContent' '429': $ref: '#/components/responses/RateLimited' '502': $ref: '#/components/responses/ChannelError' default: $ref: '#/components/responses/Error' components: parameters: ConversationId: name: conversation_id in: path required: true description: The conversation's ID. schema: $ref: '#/components/schemas/ConversationId' FlowVersion: name: Flow-Version in: header required: false description: >- The API version to use, as a date. Without it, the version pinned to your app when it was created is used. schema: type: string format: date examples: - '2026-11-01' IdempotencyKey: name: Idempotency-Key in: header required: false description: > A unique string (up to 255 characters) that makes this request safe to retry. A repeat with the same key within 24 hours returns the first answer instead of acting again. See "Idempotency" in the introduction. schema: type: string minLength: 1 maxLength: 255 schemas: TypingMessage: type: object description: > The typing indicator, sent only (`POST /v0/conversations/{conversation_id}/typing`; a `typing` piece in a webhook reply or on the stream does the same). It shows the indicator, which clears itself when your next message goes or after a few seconds. properties: kind: type: string const: typing default: typing description: >- Always `typing`. Optional on its endpoint; required where no URL names the kind. imessage: $ref: '#/components/schemas/TypingImessageOptions' ConversationAction: type: object description: The result of a typing or read call. required: - conversation - action properties: conversation: $ref: '#/components/schemas/ConversationId' action: type: string description: >- Which action the channel took. `typing` (the indicator was shown) or `read` (the read receipt was sent). enum: - typing - read x-enum-varnames: - ConversationActionTyping - ConversationActionRead ConversationId: type: string description: A conversation ID, `conv_` and a ULID. pattern: ^conv_[0-9A-HJKMNP-TV-Z]{26}$ examples: - conv_01JB8ZC3K5M7P9R1T3V5X7Z9B1 TypingImessageOptions: type: object description: | iMessage only. On another channel the block answers `422 unsupported_content` with `param` `imessage`; an unknown key `400 invalid_request`. additionalProperties: false properties: duration_seconds: type: integer minimum: 1 maximum: 60 description: >- How long to show the indicator, 1 to 60 seconds, unless a message goes first. Error: type: object description: The body of every error answer. required: - error properties: error: $ref: '#/components/schemas/ErrorBody' ErrorBody: type: object description: The details of an error. required: - type - message - hint - doc_url properties: type: $ref: '#/components/schemas/ErrorType' message: type: string description: >- A sentence for a developer saying what went wrong. Do not parse it; switch on `type`. hint: type: string description: > One sentence saying what to change to make the request succeed, specific to this case, for example "Send a template instead: POST /v0/conversations/{conversation_id}/template." For `unsupported_content` it says what to send instead over plain HTTP. Written for developers and coding agents alike. Do not parse it; it may be reworded at any time. examples: - >- Send a template instead: POST /v0/conversations/{conversation_id}/template. doc_url: type: string format: uri description: >- The documentation page for this error's type, `https://api.flow.engineer/docs/errors/`. It says what the error means, why it happens and how to fix it, with code. examples: - https://api.flow.engineer/docs/errors/outside_window param: type: string description: >- The request parameter the error is about, as a dotted path (`kind`, `telegram.is_big`, `buttons`, `cards[1].header.type`, `limit`). retry_after: type: integer minimum: 0 description: >- Seconds to wait before retrying, for errors that clear by themselves. channel: $ref: '#/components/schemas/Channel' description: >- The conversation's channel, on errors about what it takes (`unsupported_content`) and its windows (`outside_window`). conversation: $ref: '#/components/schemas/ConversationId' sender: $ref: '#/components/schemas/SenderId' channel_code: type: string description: > For `channel_error`, the channel's own error code, as given by the channel. For some other errors, Flow's own code naming the case, for example `sandbox_allowance_used` (`permission`: the sandbox allowance is used up), `sandbox_contact_limit` (`permission`: the allowance has no room for another contact), `sandbox_channel_not_included` (`permission`: the allowance does not cover this sandbox channel), `sign_in_required` (`permission`: an app made without an account cannot do this) or `sandbox_key_expired` (`authentication`: a key from `POST /v0/sandbox/keys` passed its `expires_at`; `permission`, for a send that carries no key, such as a reply in a webhook answer, from such an app). For `idempotency_conflict`, always one of `body_mismatch` (the key was used for a different request), `in_progress` (the first request is still running; `retry_after` says when to repeat it) or `secret_not_kept` (its answer carried a secret shown only once). SMS: `contact_opted_out` and `consent_required` (`permission`), `quiet_hours` (`outside_window`, with `retry_after`), and `carrier_filtered`, `unreachable`, `invalid_number`, `landline`, `registration_required`, `carrier_error` and `provider_not_configured` (`channel_error`). HIPAA accounts: `hipaa_channel_not_covered` (`permission`). Connecting your own WhatsApp number (`POST /v0/senders`): `whatsapp_token_malformed`, `whatsapp_token_rejected`, `whatsapp_app_secret_malformed` and `whatsapp_number_not_in_account` (`invalid_request`), and `whatsapp_number_taken`, `whatsapp_account_taken` and `whatsapp_number_managed` (`permission`). Simulated conversations (`permission`): `test_mode_only`, `not_simulated`, `simulated_contact` and `simulated_conversation_limit`. For a message that failed after it was queued (`message.failed`): `sandbox_contact_left` (`permission`), `window_closed` and `queued_too_long` (`outside_window`), and `channel_timeout_unknown_outcome` (`channel_error`: the channel never confirmed the send, so it may or may not have arrived). iMessage (`unsupported_content`): `group_unsupported` (a reaction, typing or a read receipt in a group conversation; `param` `kind`) and `not_on_sms` (an edit or unsend of a message that went by SMS or RCS; `param` `kind`). SMS (`unsupported_content`): `no_mms` (media to a number without MMS; `param` `media`). Edits and unsends (`unsupported_content`): `edit_window_closed` and `unsend_window_closed` (the message is older than the channel allows for the change) and `first_message_not_editable` (iMessage: the message that started the conversation). From typing and read receipts, `unknown_outcome` (`channel_error`: the call timed out after it was sent). request_id: type: string description: Flow's ID for this request. Quote it when asking for help. ErrorType: type: string description: > What went wrong. A closed list: new types arrive only with a new API version. Each type has a page at `https://api.flow.engineer/docs/errors/`, given in the error's `doc_url`; the error's `hint` says what to change for the case. - `invalid_request` (400, or 413 for a request body over the size limit): the request is malformed or a parameter is invalid. Fix the parameter named in `param`; the `hint` says what it must look like. Every endpoint answers it for a `Flow-Version` header that is not a version this API serves. - `authentication` (401): the API key is missing, malformed, unknown, revoked or expired. An `Authorization: Bearer` header with nothing after it usually means the environment variable holding the key was empty when the header was built (for an MCP client, when the server was added): the `hint` says so. Send `Authorization: Bearer fk_test_...` or `fk_live_...` with a current key; with no key at all, get a test key with `POST /v0/sandbox/keys`. A sandbox key past its `expires_at` has `channel_code` `sandbox_key_expired`: sign in to claim the app, or get a new key. - `permission` (403): the key may not do this, for example a test key using a live sender, a sandbox contact who joined another app, a new conversation from an iMessage line that may only reply, or a send past the app's sandbox allowance (`channel_code` `sandbox_allowance_used`, `sandbox_contact_limit` or `sandbox_channel_not_included`). Use the key of the right mode, have the contact send your join message, wait for the contact to message the line first, or sign in through the device flow to lift the allowance. On SMS: `contact_opted_out` (the contact texted STOP; nothing more goes to them until they text START), `consent_required` (a start to someone who never texted your numbers: record their consent with `consent` on the start). In a HIPAA account: `hipaa_channel_not_covered` (the channel has no business associate agreement; use SMS). Simulated conversations: `simulated_contact` (a real sender was pointed at a simulated contact; reply in the contact's simulated conversation), `not_simulated` (a simulated sender or test endpoint was pointed at a real contact), `test_mode_only`, `simulated_conversation_limit`. In `message.failed`: `sandbox_contact_left` (the contact joined another app on the sandbox sender while the message was queued; have them send your join message again). - `not_found` (404): no such object for this app and mode. Check the ID's prefix and that it was made with a key of the same mode (test and live data are separate). - `idempotency_conflict` (409): the idempotency key was used for a different request (`channel_code` `body_mismatch`: use a new key), or that request is still running (`in_progress`: wait `retry_after` seconds and repeat it with the same key), or it already created a secret that is shown only once (`secret_not_kept`: creating a webhook endpoint, rotating its secret, connecting a WhatsApp number; a secret that was lost must be rotated again, and a WhatsApp number connected again to see its verify token). - `outside_window` (409, or in `message.failed`): the channel will not deliver outside its conversation window. On WhatsApp the 24-hour window is closed: send a template (`POST /v0/conversations/{conversation_id}/template`, or `POST /v0/messages` with `kind` `template` to start), or wait for the contact to write. On iMessage the contact has not messaged the line (or opted in) yet, or not recently, so the failure arrives as a `message.failed` event: wait for the contact to write, then reply in that conversation. Typing and read receipts answer `409 outside_window` directly when the channel's window for them is closed (iMessage typing works only within 5 minutes of the contact's last message; WhatsApp read receipts go at any time); ignore it and send your reply. On SMS, `channel_code` `quiet_hours`: the contact has not texted in the last hour and it is inside the sender's quiet hours in their time zone; retry after `retry_after` seconds, or wait for them to text. In `message.failed`: `window_closed` (WhatsApp's 24-hour window closed while the message was queued; send a template) and `queued_too_long` (the message waited in the queue longer than its sender allows, for example while the sender was paused; send it again if it still applies). - `unsupported_content` (422): the conversation's channel does not take this kind, field, block or value, and nothing was sent; Flow never converts a message. `param` names what was refused: `kind` (the endpoint's kind is not on this channel, such as a carousel on Telegram), a field or block (`reply_to` on SMS, `telegram` on WhatsApp, `telegram.is_big`), or the field whose value is not taken (`emoji` outside Telegram's reaction emoji, `buttons` for more than 3 WhatsApp buttons). `channel` is the conversation's channel and `hint` says what to send instead over plain HTTP; each endpoint's page lists its channels. In an iMessage group conversation, a reaction, typing or a read receipt answers it with `param` `kind` and `channel_code` `group_unsupported`: leave it out. Media to an SMS number without MMS answers it with `param` `media` and `channel_code` `no_mms`. An edit or unsend of an iMessage message that went by SMS or RCS (its `network`) answers it with `channel_code` `not_on_sms`: send a new message instead. An edit or unsend past the channel's time limit answers it with `channel_code` `edit_window_closed` or `unsend_window_closed` (iMessage: edits within 15 minutes of sending, unsends within 2 minutes; Telegram: unsends within 48 hours), and an iMessage edit of the message that started the conversation with `first_message_not_editable`; `message` states the limit and how long ago the message was sent. Send a new message instead. `POST /v0/files` answers it for a type Flow does not take (programs, web pages, unknown types). A send by `file_id` of a document a contact sent before Flow malware-scanned documents answers it too: Flow sends only documents it scanned. - `new_contact_limit` (429): the sender has used its budget for starting conversations. Wait `retry_after` seconds; replies into existing conversations still go. - `sender_throttled` (429): abuse signals tripped (the same text to many new contacts, many starts with no reply, blocks), so the sender may not start conversations until `retry_after`; replies into existing conversations still go. Personalise first messages and start only conversations people expect. - `file_blocked` (422): the file failed the malware scan and was not stored. Send a different file. (When the scan cannot run at all, the upload answers a retryable `500 api_error` with `Retry-After` instead.) - `rate_limited` (429): either too many requests for this key (the per-key request limit, reported in the `RateLimit-*` headers) or sends faster than the sender's sending rate (pacing). Wait `Retry-After` (`retry_after`) seconds and retry with the same `Idempotency-Key`. - `channel_error` (502, or in `message.failed`): the channel refused or failed the message, or timed out; `channel_code` carries its own code. Read `message`, change what the channel objected to, and send again. From typing and read receipts, which call the channel at once, it is safe to ignore. SMS codes are Flow's, the same whichever carrier network carries the message: `carrier_filtered` (a carrier's spam filter blocked it: make it read like a person's message, no shortened links), `unreachable`, `invalid_number`, `landline`, `registration_required` (the number's 10DLC or toll-free registration is not approved yet), `carrier_error`, and `provider_not_configured` (the SMS sender is not fully set up on Flow's side, so nothing was sent; contact the Flow team). Flow's own codes on every channel: `channel_timeout_unknown_outcome` (in `message.failed`: the channel did not answer after the message was sent, and the resends its policy allows did not settle it, so the message may have arrived or not; check with the contact before sending it again) and `unknown_outcome` (from typing and read receipts: the call timed out after it was sent). - `not_implemented` (501): this endpoint or channel is not live yet during the beta. Use a channel that is live (Telegram, iMessage), or check the changelog. - `api_error` (500, 503): something went wrong on Flow's side. Retry with the same idempotency key after `retry_after` seconds, and quote `request_id` if it persists. enum: - invalid_request - authentication - permission - not_found - idempotency_conflict - outside_window - unsupported_content - new_contact_limit - sender_throttled - file_blocked - rate_limited - channel_error - not_implemented - api_error Channel: type: string description: >- A messaging channel. `sms` is US SMS and MMS (in pilot), from dedicated numbers only. enum: - telegram - whatsapp - imessage - sms SenderId: type: string description: A sender ID, `snd_` and a ULID. pattern: ^snd_[0-9A-HJKMNP-TV-Z]{26}$ examples: - snd_01JB8Z4Q3V6W0R2N7C5H1M9K4T responses: InvalidRequest: description: > The request is malformed or a parameter is invalid (`invalid_request`). `error.param` names the parameter. Every endpoint answers this for a `Flow-Version` header that is not a date this API serves (`param` `Flow-Version`), and every `POST` that takes an `Idempotency-Key` for a key over 255 characters (`param` `Idempotency-Key`). content: application/json: schema: $ref: '#/components/schemas/Error' example: error: type: invalid_request message: limit must be between 1 and 100. hint: >- Pass limit between 1 and 100 (default 20), and page with after or before. doc_url: https://api.flow.engineer/docs/errors/invalid_request param: limit Unauthenticated: description: >- The API key is missing, malformed, unknown, revoked or expired (`authentication`). headers: WWW-Authenticate: description: > The authentication scheme, `Bearer realm="api.flow.engineer"`. When the request carried a key that is well formed but unknown, revoked or expired, it also says `error="invalid_token"`: `Bearer realm="api.flow.engineer", error="invalid_token"`. A missing or malformed key gets the scheme alone. schema: type: string examples: - Bearer realm="api.flow.engineer", error="invalid_token" content: application/json: schema: $ref: '#/components/schemas/Error' example: error: type: authentication message: No valid API key was given. hint: >- Send the header Authorization: Bearer fk_test_... (or fk_live_...); no key yet? Get a test key with curl -X POST https://api.flow.engineer/v0/sandbox/keys doc_url: https://api.flow.engineer/docs/errors/authentication PermissionDenied: description: >- The key may not do this, for example a test key using a live sender (`permission`). content: application/json: schema: $ref: '#/components/schemas/Error' NotFound: description: No such object for this app and mode (`not_found`). content: application/json: schema: $ref: '#/components/schemas/Error' Conflict: description: > The request conflicts with the current state: the channel's window is closed (`outside_window`: WhatsApp's 24 hours for free-form messages, iMessage's 5 minutes after the contact's last message for typing), or the idempotency key was used for a different request or is still in use (`idempotency_conflict`). content: application/json: schema: $ref: '#/components/schemas/Error' example: error: type: outside_window message: >- Last message from the contact was 31h ago; WhatsApp allows only templates now. hint: >- Send a template instead: POST /v0/conversations/{conversation_id}/template. doc_url: https://api.flow.engineer/docs/errors/outside_window channel: whatsapp conversation: conv_01JB8ZC3K5M7P9R1T3V5X7Z9B1 PayloadTooLarge: description: > The request body is larger than Flow takes (`invalid_request`): 32 MB, or 101 MB for `POST /v0/files` (a 100 MB file and its multipart framing). A body that declares a larger `Content-Length` is refused before it is read. Send media as a `file_id` from `POST /v0/files` or by `url` instead of inlining it. content: application/json: schema: $ref: '#/components/schemas/Error' example: error: type: invalid_request message: The request body is larger than 33554432 bytes. hint: >- Send a smaller body; upload media with POST /v0/files and send its file_id instead of inlining it. doc_url: https://api.flow.engineer/docs/errors/invalid_request UnsupportedContent: description: > The conversation's channel does not take this kind, field, block or value, so nothing was sent (`unsupported_content`). `param` names what was refused, `channel` the channel, and `hint` what to send instead. content: application/json: schema: $ref: '#/components/schemas/Error' example: error: type: unsupported_content channel: telegram param: kind message: Carousels work on WhatsApp only. hint: >- On Telegram, send a basic message, then one buttons message with an image header per card. doc_url: https://api.flow.engineer/docs/errors/unsupported_content conversation: conv_01JB8ZC3K5M7P9R1T3V5X7Z9B1 RateLimited: description: > Too many requests (`rate_limited`): either this key's request limit (see the `RateLimit-*` headers) or the sender's sending rate (pacing) was exceeded; retry after `Retry-After` seconds (also `error.retry_after`) with the same `Idempotency-Key`. Or the sender has used its budget for starting conversations (`new_contact_limit`), or the sender is throttled after abuse signals (`sender_throttled`); wait `retry_after` seconds. headers: Retry-After: $ref: '#/components/headers/RetryAfter' RateLimit-Limit: $ref: '#/components/headers/RateLimitLimit' RateLimit-Remaining: $ref: '#/components/headers/RateLimitRemaining' RateLimit-Reset: $ref: '#/components/headers/RateLimitReset' content: application/json: schema: $ref: '#/components/schemas/Error' example: error: type: new_contact_limit message: This sender has started its 15 new conversations for today. hint: >- Retry after 3600 seconds; replies into existing conversations still go. doc_url: https://api.flow.engineer/docs/errors/new_contact_limit retry_after: 3600 sender: snd_01JB8Z4Q3V6W0R2N7C5H1M9K4T ChannelError: description: | The channel refused the call, failed or timed out (`channel_error`). `error.channel_code` carries the channel's own code where it gave one. content: application/json: schema: $ref: '#/components/schemas/Error' example: error: type: channel_error message: 'The channel refused: the request timed out.' hint: >- Typing and read receipts are safe to ignore; carry on and send your reply. doc_url: https://api.flow.engineer/docs/errors/channel_error conversation: conv_01JB8ZC3K5M7P9R1T3V5X7Z9B1 Error: description: An error. `error.type` says which (see `ErrorType`). content: application/json: schema: $ref: '#/components/schemas/Error' headers: RetryAfter: description: Seconds to wait before retrying. schema: type: integer RateLimitLimit: description: Requests allowed in the current window for this key. schema: type: integer RateLimitRemaining: description: Requests left in the current window for this key. schema: type: integer RateLimitReset: description: Seconds until the current window resets. schema: type: integer securitySchemes: bearerAuth: type: http scheme: bearer bearerFormat: fk_test_... or fk_live_... description: > An API key of one app, sent as `Authorization: Bearer `. Keys start with `fk_test_` (test mode: sandbox senders and test data only) or `fk_live_` (live mode). Keep live keys on your server; never ship them in an app or page. ```` This documentation is built and hosted on [Mintlify](https://mintlify.com), a developer documentation platform.