kapso-whatsapp

Kapso WhatsApp

Kapso (https://kapso.ai) is a WhatsApp platform vendor that fronts the Meta Cloud API. A swarm provisions one or more WhatsApp phone numbers and wires each one to either a native inbound handler (PR #560) or a workflow that dispatches a task per inbound message.

When to use MCP tools vs this skill's REST recipes

PR #560 ships thin MCP-tool wrappers for the common case only:

For ANYTHING else, drop to the REST recipes in this skill — these are the canonical reference, and the MCP tools deliberately do NOT duplicate them:

• Template messages (outside 24h window) → §"Send a template" below.
• Media (image / document / audio / video, including wide-image padding and PTT voice notes) → §"Sending media".
• Reactions (👀 / ✅ / clear) → §"Send a reaction".
• Typing indicator + mark-as-read → §"Mark as read + typing indicator".
• Signature verify (manual) → §"Webhook signature verification".
• Contact resolution → swarm user → §"Resolve a contact to a swarm user".
• Conversation history / message detail / templates list → §"Read conversation context".

If the MCP-tool send returns a 24h-window error (sessionWindowExpired: true), fall through to the template path in §"Send a template" — this is exactly what the tool's structured-error points at.

Setup

Swarm config keys (resolve with get-config key:<NAME> includeSecrets:true — Lead-only for secrets; workers should ask Lead if they need a value injected):

The curl recipes below assume $KAPSO_API_KEY, $KAPSO_API_BASE_URL, and $KAPSO_PHONE_NUMBER_ID are resolved into your shell, e.g.:

API_BASE=$(get-config KAPSO_API_BASE_URL)        # https://api.kapso.ai
API_KEY=$(get-config KAPSO_API_KEY)
PHONE_NUMBER_ID=$(get-config KAPSO_PHONE_NUMBER_ID)

The Kapso CLI is NOT installed in worker containers. Use direct HTTP or clone the gokapso/agent-skills repo for fallback scripts.

git clone --depth=1 https://github.com/gokapso/agent-skills /tmp/kapso-skills
cd /tmp/kapso-skills/skills/integrate-whatsapp && npm i  # or observe-whatsapp / automate-whatsapp

The Meta Cloud API is proxied at $KAPSO_API_BASE_URL/meta/whatsapp/v24.0/... (auth: X-API-Key). Kapso's own platform endpoints live at $KAPSO_API_BASE_URL/platform/v1/....

Inbound webhook payload (v2)

When inbound events are routed through a workflow, the workflow's webhook trigger receives whatsapp.message.* and whatsapp.conversation.* events at POST https://<your-swarm-host>/api/webhooks/<workflow-id>.

Shape (top-level keys):

{
  "message": {
    "id": "wamid.HBgL...",            // Meta message id (WAMID)
    "from": "15551234567",            // E.164 without + (sender)
    "from_user_id": "XX.00000...",    // Meta-internal user id
    "timestamp": "1700000000",        // unix seconds (string)
    "type": "text",                   // text | image | audio | video | document | sticker | location | contacts | reaction | ...
    "text": { "body": "hi" },         // only for type=text
    "context": null,                  // present when the user quote-replied another message
    "kapso": {
      "direction": "inbound|outbound",
      "status": "received|delivered|read|sent|failed",
      "processing_status": "pending|completed",
      "origin": "cloud_api",
      "has_media": false,
      "content": "hi"                 // text representation (caption / filename / body)
    }
  },
  "conversation": {
    "id": "<conversation-uuid>",
    "phone_number": "15551234567",
    "phone_number_id": "<PHONE_NUMBER_ID>",
    "contact_name": "Jane Doe",
    "status": "active",
    "last_active_at": "...",
    "created_at": "...",
    "kapso": {
      "messages_count": 10,
      "last_message_id": "wamid...",
      "last_message_text": "hi",
      "last_inbound_at": "...",
      "last_outbound_at": "..."
    }
  },
  "is_new_conversation": false,
  "phone_number_id": "<PHONE_NUMBER_ID>"
}

ALWAYS filter on message.kapso.direction == "inbound" — Kapso fires the webhook for the swarm's own outbound sends, deliveries, reads, and failures too. Only inbound events from real humans warrant a task.

Test payloads include "test": true and wamid.TEST_... ids — handle gracefully (treat as a real inbound but mark it test in your reply; do not send a real WhatsApp reply to test payloads).

Non-text message types

message.type can be text, image, audio, video, document, sticker, location, contacts, reaction, button, or interactive. Non-text inbound messages carry a type-specific object:

message.kapso.has_media is true for image/audio/video/document/sticker. message.kapso.content carries a text representation where one exists (caption, filename). message.transcript may be present for audio if Kapso pre-transcribed it.

Downloading media

Media messages carry a Meta media id (message.<type>.id), not a URL. Two-step download via the Kapso proxy:

1. Resolve the media id to a temporary URL + metadata:

   curl -s -H "X-API-Key: $API_KEY" \
     "$KAPSO_API_BASE_URL/meta/whatsapp/v24.0/<MEDIA_ID>"
   # → { "url": "https://lookaside.fbsbx.com/...", "mime_type": "...", "file_size": ..., "id": "...", "sha256": "..." }
   

2. Download the binary from that url (Meta lookaside URLs expire fast — download immediately):

   curl -sL -H "X-API-Key: $API_KEY" "<url>" -o /tmp/media.bin
   

NB: verify the exact proxy path against a real media message if your swarm has only handled text inbound so far. If the lookaside url 403s with X-API-Key, retry through $KAPSO_API_BASE_URL/meta/whatsapp/....

Recommended handling per type (proposal — adapt to what your swarm has installed)

• audio / voice notes → download the media. If the swarm has a speech-to-text capability (e.g. ElevenLabs Scribe, Whisper, or a transcription skill/tool installed on some role), transcribe via that path and feed the transcript into the conversation as if it were a text message. If no STT capability is available, download and process the raw media, or acknowledge and ask the sender to send text. Check which role owns audio/transcription in your swarm before assuming it exists.
• image → download, then describe / answer with a vision-capable model if one is available. A screenshot captioned "debug this" should get a real answer, not "I can't read images".
• document → download; read the text content (PDF/txt/etc.) and act on it.
• video → acknowledge + ask for specifics unless there is a clear transcription need.
• location → use latitude / longitude directly.
• sticker → treat as a lightweight reaction; usually no substantive reply needed.
• reaction → a user reacting to one of the swarm's messages. Usually acknowledge silently — do NOT trigger a full reply loop. (A debounce node, where present, naturally drops reaction-only events.)
• contacts → extract the shared contact info; act per the conversation.

Resolve a contact to a swarm user

The goal is to map an inbound Kapso contact (conversation.contact_name + message.from phone) to whatever user identity your swarm tracks. Two paths in order:

1. By name: resolve-user name:"<contact_name>" (fuzzy substring match). Returns the canonical profile if there's one. Useful for contacts already registered in the swarm's user registry.
2. If no match — the contact is unknown. Lead can run manage-user create name:"<contact_name>" notes:"WhatsApp +<phone>" to register them. Workers should NOT create users autonomously; ask Lead.

Always quote the phone number in manage-user notes so future lookups by phone work (until the user registry has a dedicated phone column).

Read conversation context

Use the Kapso platform endpoints via curl (no CLI needed):

API_BASE=$(get-config KAPSO_API_BASE_URL)        # https://api.kapso.ai
API_KEY=$(get-config KAPSO_API_KEY)
PHONE_NUMBER_ID=$(get-config KAPSO_PHONE_NUMBER_ID)

# List conversations for the swarm's number
curl -s -H "X-API-Key: $API_KEY" \
  "$API_BASE/platform/v1/whatsapp/conversations?phone_number_id=$PHONE_NUMBER_ID&status=active" | jq

# Get a single conversation
curl -s -H "X-API-Key: $API_KEY" \
  "$API_BASE/platform/v1/whatsapp/conversations/<conversation_id>" | jq

# List messages for a conversation — USE THE QUERY-PARAM FORM.
# (The conversation-scoped path /conversations/<id>/messages returns 404.)
curl -s -H "X-API-Key: $API_KEY" \
  "$API_BASE/platform/v1/whatsapp/messages?conversation_id=<conversation_id>&limit=20" | jq

# Single message detail
curl -s -H "X-API-Key: $API_KEY" \
  "$API_BASE/platform/v1/whatsapp/messages/<wamid>" | jq

The message list is returned newest-first.

Send a free-form text (within the 24h session window)

Per WhatsApp policy, free-form text is only allowed within 24h of the last inbound message. Outside that window you MUST use a pre-approved template (see "Send a template" below).

Common case shortcut: call the send-whatsapp-message MCP tool — it wraps exactly this REST call. The recipe below is the canonical reference and the fallback when you need fields the tool doesn't expose.

TO="15551234567"
TEXT="Hi 👋"

curl -s -X POST -H "X-API-Key: $API_KEY" -H "Content-Type: application/json" \
  -d "{
    \"messaging_product\": \"whatsapp\",
    \"recipient_type\": \"individual\",
    \"to\": \"$TO\",
    \"type\": \"text\",
    \"text\": { \"preview_url\": false, \"body\": \"$TEXT\" }
  }" \
  "$API_BASE/meta/whatsapp/v24.0/$PHONE_NUMBER_ID/messages" | jq

Returns { "messages": [{ "id": "wamid..." }] } on success. Log the wamid.

Quote-reply (thread to the original message)

Add a context object to make the message render as a reply to a specific inbound message. The reply-whatsapp-message MCP tool wraps exactly this; use the raw recipe when you need to combine quote-reply with media / templates / reactions (the tool only does text).

{
  "messaging_product": "whatsapp",
  "recipient_type": "individual",
  "to": "<phone>",
  "context": { "message_id": "<inbound_wamid>" },
  "type": "text",
  "text": { "preview_url": false, "body": "<reply>" }
}

Prefer quote-replies when answering a specific question — it keeps long conversations legible.

Sending media (image, document, audio, video)

Two-step pipeline through Kapso's Meta proxy: upload, then send by id. Sending by id is more reliable than link (no public-host requirement) — validated 2026-05-20.

1. Upload

curl -s -X POST -H "X-API-Key: $API_KEY" \
  -F "messaging_product=whatsapp" \
  -F "type=<mime>" \
  -F "file=@/path/to/file.ext;type=<mime>" \
  "$API_BASE/meta/whatsapp/v24.0/$PHONE_NUMBER_ID/media"
# → {"id":"<media-id>"}

2. Send by id

{ "type": "image",    "image":    { "id": "<id>", "caption": "..." } }
{ "type": "document", "document": { "id": "<id>", "filename": "name.ext", "caption": "..." } }
{ "type": "audio",    "audio":    { "id": "<id>" } }
{ "type": "video",    "video":    { "id": "<id>", "caption": "..." } }

Quote-reply works on media too — add "context": { "message_id": "<wamid>" } at the top level.

Wide images: pad to ~square, send as image (validated 2026-05-20)

WhatsApp scales type:image to bubble width + recompresses, so a wide 1200×630 social card renders as a tiny shrunken strip. The fix is NOT type:document — a .png sent as a document shows a plain file card with NO inline preview (must tap+download). Bad UX both ways.

Correct approach: letterbox/pad the wide image onto a ~1:1 (1080×1080) or 4:5 (1080×1350) canvas with a solid bg fill (white, or a colour sampled from the card's corner), card centered, then send THAT as type:image. WhatsApp shows ~1:1–4:5 images large WITH a preview and won't shrink them.

Pad with Pillow (ImageMagick is NOT installed in workers; pip/python3 -c with PIL works):

from PIL import Image
src = Image.open("in.png").convert("RGB"); w,h = src.size
bg = src.getpixel((0,0))                       # sample corner for fill
card = src.resize((1080, round(h*1080/w)), Image.LANCZOS)
canvas = Image.new("RGB", (1080,1080), bg)
canvas.paste(card, (0,(1080-card.height)//2)); canvas.save("out.png")

Reserve type:document for ACTUAL files — PDFs (which DO render a preview), spreadsheets, etc. — never for images.

Voice notes (PTT play bar)

For PTT (play bar in the bubble), the audio MUST be audio/ogg with Opus. MP3 sends as a generic audio attachment, no PTT bar.

ffmpeg -i in.mp3 -c:a libopus -b:a 32k -application voip out.ogg

Mark as read + typing indicator

The mark-as-read endpoint doubles as the typing indicator. POST to the same /messages endpoint with status: "read" and an optional typing_indicator:

curl -s -X POST -H "X-API-Key: $API_KEY" -H "Content-Type: application/json" \
  -d '{
    "messaging_product": "whatsapp",
    "status": "read",
    "message_id": "<inbound_wamid>",
    "typing_indicator": { "type": "text" }
  }' \
  "$API_BASE/meta/whatsapp/v24.0/$PHONE_NUMBER_ID/messages"
# → {"success":true}

The typing indicator ("typing…" dots) auto-clears after ~25 seconds OR the moment you send any message. For long-running work, re-fire this call every <25s (e.g. right before you POST your reply) to keep the dots visible. Drop the typing_indicator field to mark-as-read only.

Send a reaction

curl -s -X POST -H "X-API-Key: $API_KEY" -H "Content-Type: application/json" \
  -d '{
    "messaging_product": "whatsapp",
    "recipient_type": "individual",
    "to": "<phone>",
    "type": "reaction",
    "reaction": { "message_id": "<wamid>", "emoji": "👀" }
  }' \
  "$API_BASE/meta/whatsapp/v24.0/$PHONE_NUMBER_ID/messages"

A user can have only ONE reaction per message — sending a new emoji REPLACES the previous one (no explicit remove needed). Send "emoji": "" to clear a reaction entirely.

Send a template (outside the 24h window)

If send-whatsapp-message returns sessionWindowExpired: true, fall through to this path. WhatsApp only allows pre-approved templates outside the 24h customer-service window.

curl -s -X POST -H "X-API-Key: $API_KEY" -H "Content-Type: application/json" \
  -d '{
    "messaging_product": "whatsapp",
    "to": "<phone>",
    "type": "template",
    "template": {
      "name": "<template_name>",
      "language": { "code": "en_US" }
    }
  }' \
  "$API_BASE/meta/whatsapp/v24.0/$PHONE_NUMBER_ID/messages"

List approved templates first: GET $API_BASE/platform/v1/whatsapp/templates?phone_number_id=$PHONE_NUMBER_ID.

Webhook signature verification

Every Kapso webhook delivery includes X-Webhook-Signature: <hex> (HMAC-SHA256 of the raw body using KAPSO_WEBHOOK_HMAC_SECRET). The native handler (/api/integrations/kapso/webhook, PR #560) and the workflow webhook trigger both verify automatically — the trigger's hmacHeader is X-Webhook-Signature and hmacSecret resolves from swarm config.

To verify manually:

echo -n "$RAW_BODY" | openssl dgst -sha256 -hmac "$HMAC_SECRET" -hex | awk '{print $2}'
# Compare hex output with X-Webhook-Signature header (constant-time compare in production).

Reply etiquette

• Same language as the inbound message — match whatever the human wrote in.
• Brief. WhatsApp is not Slack — 1-3 short messages max.
• Identify yourself if it's a first interaction in the conversation, e.g. "Hi! This is the swarm's WhatsApp assistant."
• Quote-reply (context.message_id) when answering a specific question.
• If you can't help (no skill for the request, out of scope) — say so and either escalate to Lead or ask the human to use another channel.
• Always log the outbound wamid in your task output so it's traceable.

Where this fits in the swarm

Two inbound paths can exist (PR #560 adds the native one; a workflow path is the alternative):

• Native handler (/api/integrations/kapso/webhook, PR #560) — fires for any phone number registered via register-kapso-number. Verifies HMAC, dedupes by message id (KV integrations:kapso:dedupe, 24h TTL), reads the routing mapping from KV (integrations:kapso:numbers), and either dispatches a kapso-inbound task or delegates to a workflow trigger (advanced override). Also emits a kapso.message.received event on the workflow event bus.
• Workflow path — fires for unregistered numbers (or numbers whose mapping points at a workflow). A typical inbound-handling workflow chains: a react-eyes step (mark read + typing + 👀) → a debounce step (collapse rapid-fire bursts) → a gate → an agent-task triage step → a finalize step (✅/❌ reaction).

Debounce / batching: a debounce step waits a few seconds after each message and only the LAST message of a burst proceeds to the agent task — so a user firing 3 quick messages produces ONE task, not three. The agent is told the batchSize and should read trailing history and answer the whole burst in one reply. When >1 messages are collapsed, the user can be shown a "🧵 Got your N messages" note.

The agent-task triages like any other interaction; route heavier work to specialists via send-task (always include the WhatsApp source context so they can reply back).

HMAC verification is enforced (signed mode) on both paths.

Common gotchas

• Phone numbers from Kapso are E.164 without + (e.g. 15551234567). Add + when displaying to humans, drop it when calling the API.
• message.text.body is only present for type:"text". For other types read message.<type> (see the table above) or message.kapso.content for a text representation.
• Outbound status events (delivered, read) are NOT a customer interaction — skip them. Filter by message.kapso.direction == "inbound".
• Real inbound messages commonly arrive with status: "delivered" (delivered to the swarm). Do NOT skip on status — only direction signals inbound vs outbound.
• Kapso sometimes sends test payloads with "test": true and wamid.TEST_* ids. Don't reply to test payloads — just complete the task with a note.
• KAPSO_PHONE_NUMBER_ID is the swarm's own number, not the recipient's. The recipient is in message.from / conversation.phone_number.
• The message-list endpoint is /platform/v1/whatsapp/messages?conversation_id=X — the conversation-scoped /conversations/<id>/messages path 404s.
• Wide images shrink — pad them, don't send as document. type:image scales to bubble width; a wide social card becomes a strip. Sending it as type:document removes the preview entirely. Fix: pad onto a ~1:1/4:5 canvas and send as type:image. See "Sending media".
• MCP tools cover text-only. send-whatsapp-message and reply-whatsapp-message are deliberately thin — templates / media / reactions / typing / mark-as-read are NOT in the tool surface. For those, use the REST recipes in this skill directly.