What this app is
OpsPing Meeting Assistant is a Zoom Marketplace app operated by LTFI Tech, LLC ("we") as part of the hosted OpsPing incident-response service. It is not yet generally available and is pending Zoom Marketplace approval. Once approved, an authorized user can start a requested meeting from standalone Meetings, an incident's Start Meeting Assistant panel, or Slack. A hosted Go backend checks that the connected Zoom user owns the meeting and starts Real-Time Media Streams (RTMS); a single-purpose Python Fargate recorder captures audio without a bot participant or Meeting SDK. Transcripts and AI summaries are available in Meetings, with optional incident and Slack delivery. Four application OAuth scopes are requested: user:read:user, meeting:read:meeting, meeting:read:meeting_audio, and meeting:update:participant_rtms_app_status. Zoom also auto-adds rtms:read:rtms_started and rtms:read:rtms_stopped for lifecycle events. Scopes authorize API capabilities; they do not mechanically restrict access to a single requested meeting. OpsPing enforces the initial owner check and correlates callbacks to the expected meeting occurrence and stream.
Data we collect
- Meeting audio — received through Zoom RTMS after the meeting owner starts the stream and participants see Zoom's native RTMS disclosure. A single, isolated AWS Fargate task (us-east-1) writes one 16 kHz mono WAV per participant and uploads them to the session's private staging prefix in Amazon S3, encrypted at rest with AWS KMS (SSE-KMS). After the meeting ends, the audio is transcribed by Cloudflare Workers AI; the staged recordings are deleted once the transcript is stored, unless your organization enables the keep-audio setting, in which case they stay under the session's private S3 prefix and follow the organization's retention window.
- Participant display names and IDs — participant identifiers and display names delivered in RTMS events, used solely to label speakers in the transcript.
- The transcript — the text produced from the meeting audio.
- The AI-generated summary — produced from the transcript and available in Meetings, with optional delivery to your incident timeline and/or Slack thread.
- Summary inputs — audio-derived transcript text with the meeting ID prepended (sent in chunks, with partial summaries on rolling/combine steps) and participant display names/speaker labels embedded in it, organization vocabulary entries (canonical term, definition, and "often misheard as" aliases), where canonical terms are transcription hotwords and terms/definitions/aliases feed summary correction, summary instructions, draft and partial summary context, and — for the terminology-correction pass when the session is linked to an incident — linked incident fields (title, description, responders, impacted services, timeline, postmortem action items; assistant-posted timeline entries excluded; capped at 20,000 characters), used to produce the summary. These are sent to the
/summarizeroute of the same Worker; audio is not. - Operational metadata — meeting ID and UUID, RTMS stream ID, timestamps (requested, stream started, ended), session status, the authorizing Zoom user's user ID, and RTMS connection details (signaling server URLs and operator), plus task CloudWatch logs (participant display names and IDs, meeting timings), used to run and audit the feature.
- Zoom tokens — where a user connects their Zoom account (via the Connect Zoom button Slack offers when a launch is triggered, or in the OpsPing admin console at Settings → Integrations), the resulting OAuth tokens are encrypted and used only to verify that the user is the meeting owner, to start/stop the RTMS stream, and to process deauthorization requests. The app requests
user:read:userto bind the connection to your Zoom user ID; the Zoom email is not stored.
Where data is processed
Zoom RTMS audio is captured by an ephemeral AWS Fargate task in the us-east-1 region: a single-use container that opens the Zoom RTMS signaling and media connections, writes one 16 kHz mono WAV per participant, uploads each recording to the session's private staging prefix in Amazon S3 (SSE-KMS), and is destroyed afterward. After the meeting ends the OpsPing backend splits the staged recordings into short segments and sends them to the /transcribe route of OpsPing's opsping-meeting-summary Worker, which transcribes them with Cloudflare Workers AI (@cf/openai/whisper-large-v3-turbo); audio is processed on Cloudflare's global network, outside the United States and outside us-east-1, and the staged recordings are deleted once the transcript is stored unless your organization enables keep-audio. Transcripts are stored in a private Amazon S3 bucket in us-east-1 encrypted with AWS KMS (SSE-KMS); summaries and meeting metadata are stored in the OpsPing PostgreSQL database in us-east-2. Transcript text (not audio) plus the summary inputs above are sent to the /summarize route of OpsPing's opsping-meeting-summary Worker, which calls Cloudflare Workers AI; Cloudflare-hosted inference may process transcript text on Cloudflare's global network, outside the United States and outside us-east-1. Traffic to Zoom, OpsPing, S3, and Cloudflare is encrypted in transit (TLS).
AI summaries
Transcripts are summarized by Cloudflare Workers AI (the default provider operated for LTFI Tech, LLC — a Cloudflare-hosted model; transcript and summary text may be processed on Cloudflare's global network). Organization admins select default (Workers AI only), customer (the organization's own HTTPS AI endpoint for summaries; transcription always uses Cloudflare Workers AI and is unaffected by this setting), or both. For summarization, Workers AI receives transcript text with the meeting ID prepended (including participant display names/speaker labels embedded in the transcript) plus prompts/instructions (including glossary entries — canonical term, definition, aliases — and summary instructions), draft and partial summary context, and linked incident fields (title, description, responders, impacted services, timeline, postmortem action items). We do not send saved revisions, correction suggestions/improvement-queue items, or Slack channel/thread routing. On this default path, Cloudflare is the only AI inference subprocessor: per Cloudflare's Workers AI Data usage documentation, standard inference requests do not retain inputs or outputs after they are needed to deliver the response, and Cloudflare does not use Customer Content processed by Workers AI to train models. Those commitments are subject to the exceptions described there: stateful features that store content (we enable none for the meeting Worker), operational metadata such as account and model identifiers, timestamps, request and response sizes, and token counts, and safety, security, or legal exceptions. We do not enable AI Gateway logging, Worker request logging, or Workers observability for the meeting Worker. For each summary OpsPing records the provider (default or customer_endpoint) on the summary record and writes a meetingbot.summarized audit entry listing the provider(s) whose summaries were delivered. In customer or both mode, the organization's own AI endpoint (using the endpoint URL and prompt the organization configures) also receives the same prompt, transcript chunks, partial/draft context, and linked-incident correction content; its retention and training terms apply. We do not sell, rent, or trade meeting data. We share transcript text with the meeting ID prepended, and the summary inputs described above, with Cloudflare Workers AI for inference and deliver summaries to the incident record and the Slack channels the organization configures, and no meeting content is used to train models on the default Cloudflare path. Meeting audio is sent only to Cloudflare Workers AI for transcription, never to a customer-configured endpoint.
Retention
Transcripts and summaries are deleted by the daily retention job when the organization's retention setting (1–365 days; default 30 days) elapses. The daily job deletes each expired transcript object from S3 before deleting its row; when that object deletion fails, the row is kept and retried on the next daily sweep. The transcript bucket's separate S3 lifecycle rule expires objects after its configured period (fixed 400-day fail-safe: 365-day maximum tenant retention + 35-day margin) as a backstop. Saved summary corrections (revisions) and correction suggestions / improvement-queue items carry meeting content and follow the same schedule. This setting (and the rest of the meeting-assistant configuration — summary mode, customer AI endpoint, Slack posting) is administered by the organization's OpsPing admins in the OpsPing admin console at Settings → Integrations. Meeting audio is staged encrypted in the session's S3 prefix for transcription and is deleted as soon as the transcript is stored; the exception is an organization that enables keep-audio, which keeps the staged per-participant WAVs under the session's S3 prefix until the transcript reaches its retention deadline. Operational metadata is kept only as long as needed to operate the feature and honor the same retention window. Copies of summaries already posted to Slack channels or incident timelines are not deleted by this window — Slack copies follow the workspace's retention, and internal incident-timeline copies follow the incident record's and organization's lifecycle.
Host and participant controls
Nothing joins the meeting; capture runs through Zoom RTMS, which is visible and controllable:
- Only the meeting owner who connected the Zoom account can start capture in OpsPing; a launch by any other connected user is rejected before anything is started.
- Zoom provides native RTMS sharing disclosures and controls; a sharing notice may appear instead of an approval prompt. Notices do not guarantee explicit all-participant consent and must not be relied on to prevent capture. The host and customer must obtain any legally required participant consent before starting.
- The owner must join before starting. Zoom's default continuous meeting sharing allows the app to continue after the owner transfers host and leaves; the new host can control media access. End Meeting for All ends the meeting and capture.
- Use the RTMS sharing controls available in your Zoom client to manage sharing; do not rely on dismissing a disclosure to prevent capture.
- Organization admins can disable the assistant, change retention, or turn on keep-audio in the OpsPing console at Settings → Integrations.
Deletion
You stay in control of your data:
- Zoom deauthorization. Removing the app from Zoom sends an authenticated deauthorization event; disconnecting in OpsPing at Settings → Integrations initiates the same connection cleanup. OpsPing revokes connection authority and fences new work, confirms recorder task stops before deleting S3 content, and deletes meeting-assistant database records last. Failures retain durable cleanup state for retry; deletion is not promised instantaneous. The 400-day S3 lifecycle ceiling is a backstop. Copies of summaries already posted to Slack channels or internal incident timelines are not recalled — Slack copies follow workspace retention, and incident copies follow the incident record's and organization's lifecycle.
- Organization deletion. Organizations are deleted by OpsPing support (there is no self-serve organization deletion). When an organization is deleted, the meeting-assistant database rows — sessions, transcripts, summaries, saved corrections, correction suggestions, and Zoom connections — are deleted and the stored transcript objects and staged audio recordings under the organization's prefix are deleted; individual object-deletion failures are logged and the transcript bucket's S3 lifecycle rule above removes any object left behind. The org's vocabulary glossary and summary instructions are organization configuration used to bias transcription and summaries; they may include admin-approved short excerpts from corrections, are kept until the org's admins delete them, and go with the organization when it is deleted, outside the meeting-content retention window.
- Manual requests. To request deletion of your data at any other time, email us below and we will comply within 30 days.
Security
Zoom tokens are encrypted at rest (AES-256-GCM). RTMS processing tasks are single-purpose, accept no inbound network connections (their security group denies all ingress), and receive only their per-session HMAC-signed token, a per-session join config containing non-secret RTMS coordinates plus a signature computed server-side, and short-lived presigned upload URLs for the session's own audio staging prefix (no general S3 or KMS permissions) — no long-lived Zoom OAuth tokens, no Zoom client secret, and no general cloud credentials. Zoom webhooks (deauthorization and RTMS events) are signature-verified before any data is changed. Transcripts and staged audio live in a private, KMS-encrypted S3 bucket in us-east-1. Our general practices are described at opsping.com/security.
Changes
Material changes to this policy will be posted on this page with an updated effective date.
Contact
LTFI Tech, LLC · support@opsping.com