Skip to main content
A sub-processor is a third party that processes data on our behalf. This is the complete list for the platform as it stands.
We will give notice before adding a sub-processor that receives customer data. If you need the commitment in contract form, ask — it belongs in the DPA rather than on a page we can edit.

Message content

One sub-processor receives anything a person typed or wrote: TypeSafe, for every message. Mail to and from an agent’s email address stays between your own AgentMail account and us; see Your own tools.
Message classification
required
What reaches them: the text of a message addressed to your agent, and the word slack, linear or email. Capped at 4,000 characters. No identifiers of any kind. Full detail.Why: to decide whether an inbound message is work, so that a greeting does not open a billable case and a real request is never silently dropped.Model: a pinned version, so answers cannot change under us without a deliberate change on our side.Where they run: the United States, on AWS with inference on Modal, Nebius and CoreWeave — who process prompts without storing them. Our own control plane is in India (ap-south-1), so classification is the one hop that leaves the region.Their commitments: TypeSafe state that their model is not trained on customer requests or responses. See Retention for what we have and have not contracted for.

Live view

Only when a person opens an agent’s computer live.
Live view relay (Realtime TURN)
What reaches them: the encrypted stream between the agent’s box and the viewer’s browser: the screen, and the viewer’s input while they hold control. It is encrypted end to end (WebRTC, DTLS-SRTP) and Cloudflare cannot read it. They also see the IP addresses and ports of the box and of each viewer, and the timing and volume of the traffic. No case content, no names, no workspace. Full detail.Why: so a person can watch an agent’s computer and take control of it, without either end learning the other’s address and without the screen passing through our servers.Where they run: Cloudflare’s global network. Traffic is relayed by the location nearest each end, which may be outside India.When it is used: only while somebody watches. Nothing is sent when nobody does. Each viewer’s credential lasts 15 minutes and is revoked when they leave.Their commitments: how long Cloudflare keeps connection metadata is being confirmed against their terms, and we will not state it here until it is.

Platform infrastructure

These receive operational data — accounts, case records, telemetry, case files, payments, the waitlist — and none of them receive your repository; a file from it reaches AWS only when the agent publishes it to a case.

Your own tools

The last row is not our sub-processors: they are your vendors, under your own agreements with them. AgentMail is one of them once an admin connects your workspace’s own AgentMail account at Integrations. We hold its key sealed and use it only to set up your mail domains and each agent’s inbox, and each agent’s own key, which reaches only its inbox. Your agents’ mailboxes live in your AgentMail account; mail passes between it and our API for the agent’s work, the way Slack messages do. Mail AgentMail categorises as spam, blocked or unauthenticated is never delivered to an agent, message bodies are treated as untrusted input, and our audit records hold metadata, never bodies. When an email body is classified it is sent to TypeSafe as described above, like a Slack message. Traces go to whichever target you chose at enrollment — ours or your own self-hosted instance. If you chose your own, no trace data reaches us at all. See Where things run.

Who receives nothing

Worth saying explicitly, because it is the question behind the question:
  • No model provider receives your source code from us. Prompts containing your code go from your machine to your provider on your key. We are not in that path and could not insert ourselves into it.
  • No sub-processor holds a credential to your repositories, or a readable key to a tool you connect. Each agent’s GitHub App key, and the secret fields of every connected tool, are sealed in our own vault, in our AWS account, with AWS KMS. They are used only to mint short-lived tokens, or to hand a tool’s keys to your runner sealed to it (how).
  • Cloudflare cannot read the live view. It relays an encrypted stream and holds no key to it.
  • Nango receives nothing. The OAuth layer that manages your Slack, Linear and GitHub connections is the self-hosted Nango server, running inside our infrastructure with its own database and encryption key. Your connection tokens never reach Nango, Inc., which is why the company is not on the list above.

If you are filling in a questionnaire

The two answers people usually need:
  • Data residency: the control plane runs in AWS ap-south-1. Message classification is processed in the United States and cannot be placed elsewhere today. An agent’s mailbox lives wherever your own AgentMail account keeps it. The live view’s encrypted stream is relayed by Cloudflare’s location nearest each end, which may be outside India. If your procurement requires a specific region, tell us before you sign rather than after — it is a real constraint on where we can put your tenant, not a checkbox.
  • Sub-processor change notice: see the note at the top of this page, and get it in the DPA.