Privacy policy
This text describes how Pagr is built and what it stores. Last updated 2026-08-24.
This policy explains what Pagr (“we”) collects when you use the service, why, and how long we keep it. The operator is Stagberry Labs LLC. Questions: support@pagr.dev.
The short version
- We store your phone number, recent iMessage content (30 days by default), and session metadata.
- We do not store your source code, your Claude or Codex credentials, or full agent transcripts.
- You can delete your account and data from the dashboard.
What we store
| Data | Why | Retention |
|---|---|---|
| Account identity: sign-in ID and your verified phone number (E.164) | To know which phone is allowed to control your agents | Life of account |
| iMessage content you send and the replies we send you | Conversation continuity: “that one”, “the second option”, follow-ups | 30 days by default (configurable), then deleted |
| Compact conversation summaries | So Pagr remembers context without keeping every message | Until superseded or account deletion |
| Session metadata: provider, status, task summary, timestamps, opaque project IDs and display names | To route messages to the right agent and tell you what is running | 90 days after the session ends |
| Normalized agent events: started, needs input, approval requested, completed, failed, with short previews | To notify you and to bind approvals to exactly what you saw | 90 days |
| Device metadata: device ID, public key, last seen, bridge version | Pairing, signing, revocation | Until you revoke the device |
| Preferences: notification policy, approval policy, default reviewer, project aliases | To behave the way you asked | Life of account |
| Screenshots you attach | To forward them to the agent on your Mac | Deleted within 24 hours |
| Billing mirror: subscription status and plan (payments handled by Stripe) | Entitlements | As required for tax and accounting |
| Audit and security records: who approved what, when, from which device (hashes and previews, not payloads) | Security and abuse investigation | At least 90 days |
What we never store
- Full repository contents or source files
- Your Claude, ChatGPT, or Codex credentials or API keys
- Contents of
.envfiles or other secrets on your Mac - Full Claude or Codex transcripts or raw terminal logs
- Full local filesystem paths (the bridge maps them to opaque IDs)
Our hosted service is designed so that it does not need a copy of your repository. The open-source bridge on your Mac talks to the coding runtimes locally and sends us typed events with short, redacted previews. Note that Claude Code and Codex themselves send code to Anthropic and OpenAI to do their work, under those providers’ own terms and privacy policies — that is outside Pagr’s control.
Who we share data with
- An iMessage delivery provider, to send and receive your texts.
- An authentication provider, to verify your phone number and sign you in.
- Stripe, for payments. We never see your card number.
- An LLM provider (currently OpenAI) that processes your messages to understand intent. Message content is sent for processing and is subject to that provider’s API data terms; we configure it not to train on your data where the provider offers that option.
- Error-tracking and tracing services, with message content, paths, phone numbers, and secrets masked before they leave our servers.
We do not sell personal data and do not use it for advertising.
Connecting Pagr to an AI assistant (MCP)
Pagr can be connected to an AI assistant — Claude, ChatGPT, Codex, or any other MCP client — as a remote MCP server. This section describes that path specifically, because the data flows differently from texting.
- What the connection grants. A connected client holds an access token scoped to some combination of
read(see your Macs, projects, sessions and pending approvals),control(start, steer and stop agent sessions), andapprove(allow or deny what an agent is waiting on). You choose the scopes on the consent screen and can revoke the token at any time from Security in the dashboard. - What we send to the client. Only what the tool you called returns: device names and online state, project display names and aliases, session status and short task summaries, approval previews and risk tiers. Filesystem paths, preview hashes, provider request and session identifiers, public keys and repository hints are stripped before anything leaves.
- What we never send. Source code, file contents, full agent transcripts, terminal output, your iMessage history, your phone number, or your agent credentials. There is no tool that reads a file or runs an arbitrary command.
- What the assistant's operator sees. Everything the tool returned, because it passes through their model to answer you. Anthropic, OpenAI, or whoever operates the client you connected handles that under their own terms and privacy policy, not ours.
- What we record. Every tool call and every refused tool call is written to your audit log with the tool name, the outcome, the token used and the fact that it came from MCP rather than from a text. Tool arguments are not stored.
- Demo connections. If we have issued a reviewer demo token, requests using it return fixed sample data and touch no real account. Nothing about a demo session is associated with a person.
Pagr is not affiliated with, endorsed by, or a partner of Anthropic or OpenAI. Connecting Pagr to their products does not change that.
Analytics
We measure the activation funnel (signed up, paired, first message, and so on) with pseudonymous IDs and event names only. Analytics never receives message text, source code, repository paths, or secrets.
Your controls
- Revoke any device from the dashboard; the bridge disconnects immediately.
- Unlink your phone number.
- Change message retention in settings.
- Text STOP at any time to stop all messages.
- Delete your account. This revokes devices, unlinks messaging, deletes attachments immediately, and deletes messages, sessions, projects, and workflows after any retention required for billing or legal reasons. It does not touch your local repositories or agent sessions.
Security
See the security model for how device keys, signing, and approvals work. Report vulnerabilities to security@pagr.dev.
Children
Pagr is for developers and is not directed at anyone under 16.
Changes
We will post changes here and, for material changes, message you before they take effect.