An AI agent you can trust with your customers
The agent checks who it's talking to, asks before it touches money and shows its work on every reply. You set the limits, and you can switch it off at any time.
Built so a wrong step is hard to take
These rules are part of how the agent works. They apply on every plan.
Identity first
The agent only looks up or changes things for a customer it has verified.
- A signed identity check in the website chat
- The sender address on email
- A one-time code by email before anything sensitive
- Identity fields are filled in by code. The model never types them.
Risk levels and approvals
You decide what the agent may do on its own.
- Safe actions run on their own
- Some actions wait for the customer to confirm in the chat
- Money and account changes wait for a person on your team to approve
Limits on every action
Even an approved action has guardrails.
- Amount limits, such as refunds up to $50
- Daily caps per action
- A unique key on every action, so a retry can't run it twice
- Every action is logged on the conversation
Personal data stays out of logs
Names, emails, card numbers and other personal details are masked before traces and logs are saved.
Screening for prompt attacks
Every customer message is checked for attempts to trick the agent before the agent acts on it.
Keys stored encrypted
Sign-ins for your apps and the agent's own keys are stored encrypted. They're never saved as plain text.
Teams kept apart
Each team's data is kept separate on every row, so one team's agent can only see its own conversations, knowledge and apps.
See everything and stop anything
Kill switches
Turn the agent off for one inbox, for your whole team or everywhere, in one click.
Explain any reply
See every step behind a reply: the rules it followed, the sources it read and the tools it called.
Spending caps
Set a monthly limit. When it's reached, the agent hands new conversations to your team.
Tests before going live
Replay your past conversations against the agent before it talks to a real customer.
Draft mode
Start with the agent writing drafts that your team sends. Switch to Auto when you trust it.
Self-host the inbox
Run the open source inbox on your own servers to keep conversations on your own hardware.
Read the code behind your inbox
The inbox, channels, help center and reports are open source under the MIT license. You can review how they handle your data, or run them on your own servers.
# Found a problem? Tell us first. to: security@magpiereply.com include: what you found, how to repeat it we reply: as soon as we can # Please don't share it in public until it's fixed.
Security questions
Do you train AI models on our data?
We don't use your conversations to train AI models. The agent uses Claude through Anthropic's API, and Anthropic's commercial terms say API inputs aren't used to train their models.
Where is our data stored?
MagpieReply Cloud runs on managed Postgres. If you self-host, everything stays on your own servers.
Can we delete our data?
Yes. Delete your account in the app, or email us and we'll delete it for you.
Can we read the code?
Yes. The inbox is open source under the MIT license. Read it on GitHub.