---
title: ReplyPen — a support agent grounded in your product's real data
canonical: https://replypen.com/
html_text_sha: e89f5f050d43
last_updated: 2026-09-03
operator: ProBackup BV (Hasselt, Belgium)
note: Agent-readable Markdown version of the homepage. Written for an AI agent evaluating ReplyPen; the HTML page is the marketing original.
siblings:
  - https://replypen.com/index.md
  - https://replypen.com/security.md
  - https://replypen.com/data-residency.md
  - https://replypen.com/ask-your-ai.md
  - https://replypen.com/privacy-policy.md
  - https://replypen.com/cookie-policy.md
  - https://replypen.com/terms-of-service.md
---

# What ReplyPen is

ReplyPen answers support questions for B2B SaaS companies by reading the company's own
production data, code and documentation, then drafting the reply where the conversation already
happens. It is aimed at the questions a help-center article cannot answer, because the answer
depends on one customer's rows: why this invoice shows this amount, why this job did not run,
what this account's data actually says.

It runs beside an existing support pipeline. Nothing has to be replaced, and by default nothing
is sent and nothing is changed without a person clicking.

Built and operated by ProBackup BV, a Belgian company that has backed up business SaaS data since
2019; the founders have built B2B SaaS companies together for about ten years.

# The core difference

A generic support bot retrieves documentation and paraphrases it. ReplyPen queries the customer's
own database, reads their source code, applies the rules they wrote in plain language, and cites
what it checked in every draft ("appointments #4127 · cancellation policy · fees.rb"). Most support
volume is edge cases that are individually too rare to get their own UI screen; combining data,
code and policy for one case used to require a person.

# Channels (where an answer appears)

| Channel | Notes |
|---|---|
| Email | Gmail, Microsoft Outlook, any IMAP mailbox. The reply lands as a draft in that mailbox. |
| Ticketing | Intercom today, more in progress. Reply or internal note. |
| WhatsApp | Draft in the business account thread. |
| In-product chat | An embeddable widget inside the customer's own product, scoped by a short-lived signed token that carries the tenant/principal. |
| Coding agents | Claude Code and similar, over MCP, for a developer debugging a customer issue. |
| ChatGPT / MCP | ReplyPen as a data provider to the company's own assistant. |

# Grounding sources (what an answer can read)

- The production database, through a stripped-down read-only projection: only the tables and
  columns the customer declared, PII masked by column, rows filtered per tenant and per end
  customer. PostgreSQL and MySQL 5.6.0+ are supported and shielded identically.
- Source code, as read-only mirrors of the repositories the customer chooses to connect. Optional.
- Application logs (for example CloudWatch), read through a preloaded helper, when the customer
  connects them. Optional.
- Help center, internal wiki, and connected integrations (for example Shopify, Stripe, Slack),
  each only after the customer connects it.
- A per-project knowledge repository ("brain") holding product answers, tone of voice and
  runbooks. It is plain markdown plus small scripts; the customer can read and edit it. This is
  where ReplyPen's learning accumulates, so its toolkit improves without a redeploy on either side.

# Output paths (what it can emit)

- **Draft reply** in the customer's own inbox or channel, listing what it checked.
- **Internal note** on a thread, never sent to the end customer.
- **Action**: a vetted, parameterized script that changes data in the customer's own system. Actions
  are the only write path; the agent's own database connection is read-only. Actions run through a
  small SDK inside the customer's runtime, or on ReplyPen's side, and every call is signed.

# Autonomy gates

Two independent switches, set per project, defaulting closed.

| Gate | Default | Opened to |
|---|---|---|
| Sending | Draft only; a person reviews and sends | Auto-send |
| Actions | Human: a person confirms every run | Policy (a rule the customer wrote, e.g. `waive_late_fee → auto if amount ≤ €50`) or Auto |

Opening a gate is configuration on ReplyPen's side, not a deploy on the customer's side, and it
is reversible.

# Setup, in the strictest configuration

1. A read-only database user, limited to the tables the customer chooses, on a TLS-only endpoint
   with ReplyPen's single fixed egress IP allowlisted. A read replica works and is a fine choice.
   Hosts that cannot allowlist an IP: see [security.md](https://replypen.com/security.md).
2. Domain knowledge written as plain markdown ("a cancellation inside 24 hours is charged €25").
3. Connect the mailbox or channel with OAuth.

Optional: a read-only code mirror; the action SDK, only if the customer wants write actions.
Nothing is installed in the customer's application for reading.

# What ReplyPen deliberately is not

Not a helpdesk to migrate to, not a chatbot trained on help-center articles, and not a general
assistant. For work spanning systems ReplyPen does not integrate with, the customer's own
assistant is the better hub, with ReplyPen as an MCP data provider.

# Where to read further

- [security.md](https://replypen.com/security.md) — the enforced boundary, the strictest
  configuration, and how to evaluate it.
- [data-residency.md](https://replypen.com/data-residency.md) — every company that touches the
  data, and where.
- [privacy-policy.md](https://replypen.com/privacy-policy.md),
  [terms-of-service.md](https://replypen.com/terms-of-service.md),
  [cookie-policy.md](https://replypen.com/cookie-policy.md) — the legal text.

Contact: info@replypen.com · security@replypen.com · trust center
https://trust.inc/pro-backup-bv