BonkerBot

Multi-signal, with a corroboration floor

Every check contributes a bounded signal. No single one can cross an action threshold alone — that always takes a second, independent signal agreeing. Keyword-score bots have no such floor.

How the scoring works

🖼️

It checks the text inside images

Screenshot scams work because every keyword filter on the platform goes blind at the image boundary. OCR pulls the text back out, and perceptual hashing still matches a re-cropped copy of a known scam image.

🧊

A filter, not a reader

Messages that don't match are dropped in memory — not stored, not logged, not attached to your account. Nothing is sent to any third-party API. It cannot see your DMs with other people; no bot on Discord can.

What the bot can see

📊

Reputation is graduated, not binary

A global 0.0–10.0 score follows an account across every protected server. Velocity-capped, passively recovering, and extra-penalised when three independent servers flag the same user.

⚖️

Due process is built in

Appeals, undo buttons on every automated action, human mod review with Confirm/Dismiss/Escalate, a full audit log, an emergency kill switch, and user-initiated data export and deletion.

🔓

Open where it counts

Every line that governs what happens to a person — scoring, thresholds, trust maths, appeals, retention — is public and auditable. The scam pattern list is not, because a published rule set is a bypass checklist.

How the project is split

Two questions people ask first

"Is it reading everything I type?" It's a sieve, not a reader. Messages are checked in memory against known scam structures; anything that doesn't match is dropped without being stored, logged, or counted against you. Storage happens after the flag threshold, not before it — so the overwhelming majority of messages leave no trace at all. It has no memory of what you talk about, and it physically cannot see your DMs with other people, because Discord's API doesn't expose them to any bot. → What the bot can see

"If it's open source, can't scammers just read the rules?" They could, which is why the rules aren't the open part. How the bot treats people is public — the scoring, the thresholds, the trust maths, the appeals, every path that can punish someone. What it looks for stays private, because publishing a detection list hands scammers a free QA suite for their next campaign. → How the project is split

The honest version

BonkerBot is not an anti-nuke bot. It won't stop a rogue admin deleting your channel list, it has no server backup or restore, and it has no CAPTCHA gate. If that's your threat model, run Wick or Security Bot, plus Beemo for raid floods. Those bots defend the perimeter. BonkerBot defends the conversation. They don't overlap enough to conflict — most servers past a few thousand members want both.

It's also young. Wick reports ~837,000 servers; BonkerBot is at v1.0.0 with no production track record to speak of. The claims on this page are architectural — the design refuses to act on a single weak signal — not empirical.

The full, unflattering comparison lives in Why BonkerBot, including a section on where it's genuinely worse than the alternatives.

Getting started

/setup-server run

That's the guided flow — it walks you through the mod review channel, moderator roles, and which detection modules to enable. Everything it sets can be changed later with /botconfig.

New here? Start with the overview.

Want the full command list? Command reference.

Running your own instance? Self-hosting — note that a self-hosted bot brings its own patterns and is not on our trust network.

BonkerBotCatches the scam, not the keyword.