Carl-bot review: what it is good at and where it runs out

We make an AI moderation tool, so treat this as an interested party's assessment. Every claim about Carl-bot below comes from Carl-bot's own documentation, and the rating comes from a public page you can open yourself. Where we could not verify something, we say so instead of guessing.

What Carl-bot is

Carl-bot is a general purpose Discord bot best known for reaction roles, with a substantial moderation side alongside logging, embeds, feeds, greetings, starboard, suggestions, tags and triggers. It is configured mostly through a web dashboard, and its own documentation says it is "highly recommended to use the Dashboard for setting up automod".

What Carl-bot does well

The punishment set is the richest of any bot we have looked at. Automod can respond with delete, warn, tempmute, mute, timeout, kick, tempban, ban, message, DM, or defer. Durations are written naturally as 3h42m, and crucially you can combine punishments by separating them with commas, so delete, tempmute 20m is a single valid response. Most bots give you one action per rule. Carl-bot gives you a policy.

Defer is a genuinely original idea. Instead of acting automatically, the defer punishment sends the offending context to a drama channel where moderators vote with reactions. That turns the ambiguous cases into a queue for humans rather than forcing you to choose between over-blocking and under-blocking. It is a premium feature, and it is the most thoughtful design in the category.

The module coverage is broad. Automod covers message spam, attachment spam, bad words, caps limit, invites, links, mentions, and a honeypot. Each has its own rate limits and its own punishments.

The rate limit model is honest about needing configuration. Message spam "will not be active without setting a rate limit of at least 1+ messages in 1+ seconds first", and the docs warn that once active, the triggering message is deleted even if delete is not among the punishments. That is the kind of side effect most products let you discover in production.

Whitelisting and scoping are well thought out. Roles and channels can be whitelisted so automod ignores them, channels can be set media-only, and there is a deletefiles toggle for unsafe file types with the safe list spelled out: png, jpg, jpeg, gif, svg, bmp, tif, webp, webm, mp4, mov, pdf, txt, mp3, flac and wav.

Warn thresholds are persistent by design. Warns do not automatically expire, and the threshold punishment triggers on every new warning while a member is above the limit, which is a deliberate and defensible choice rather than an oversight.

Where Carl-bot runs out

Every automod module is a string match, a rate limit, or a trap. That is the ceiling.

Bad words only catches what you wrote down. Not the leetspeak version, not the homoglyph version, not the one with a zero-width character in the middle, and not the same idea in another language. The usual evasions apply: fr33 n1tr0, h8, 5c4m, h a t e, s.c.a.m.

Rate limits measure volume, never meaning. Message spam, attachment spam, caps and mentions all ask how many and how fast. A targeted threat, a grooming attempt or a convincing scam is one calm, correctly spelled message that trips none of them.

The honeypot catches bots, not people. It is an excellent trap for automated spam accounts, and a member deliberately making one person's life miserable will never post in it.

There are no harm categories. Nothing identifies harassment, hate, threats or self-harm as categories in their own right, so each becomes a censor list you maintain.

Nothing weighs conversation context. Each module judges one message alone.

The documentation describes no image or video content analysis. Attachments are rate limited and unsafe file types can be deleted, but what is inside a picture is not examined.

What users say

Two things about that number. top.gg rates on a 0 to 100 scale rather than out of five. And it is lower than the other bots we checked on the same scale and the same day, where Dyno was 87, Wick 84 and Sapphire 96.

We are reporting it because leaving it out would be dishonest, and we are not going to build a narrative on it, because we cannot tell you why it is what it is. We could not systematically sample the reviews behind it from sources we can cite, so we do not know whether it reflects the moderation features, an outage, a pricing change, or something else entirely. A rating is not an explanation. Carl-bot remains one of the most widely used bots on Discord and the feature work described above is genuinely good.

Where Supervisor fills the gaps, and where it does not

Supervisor reads what a message means rather than matching strings and counting events:

• No censor list to spell around, so leetspeak, homoglyphs and invisible characters are read as what they mean.
• 16 labels including harassment, hate, threats, scams, self-harm and illegal activity.
• Conversation context weighed when a message is ambiguous alone.
• Images and video moderated on content, not just rate limited by count.
• Over 100 languages without a list per language.

Where it does not fill the gap, plainly:

• Supervisor cannot ban, kick, mute or tempban. Its actions are Delete, Timeout and Warn. Carl-bot's punishment set is far richer, and there is no Supervisor equivalent of combining several punishments in one rule.
• Supervisor has nothing like defer. There is no moderator voting queue.
• Supervisor has no honeypot, no persistent warn threshold system, and no rate limits.
• Supervisor has no reaction roles, embeds, feeds, starboard, tags or triggers. Three slash commands, and everything else on the dashboard.
• Supervisor is paid, at £13.99 a month or about £4.99 a month on the three year cycle, with a seven day trial and £0.25 of moderation free on signup.

There is a full comparison of the two approaches if you want the mechanism side in depth.

Who Carl-bot is right for

Carl-bot is right for you if you want reaction roles done properly, a genuinely flexible punishment system, and automod modules you can tune precisely. The combination of composable punishments and the defer queue makes it the best choice in this category for a server that wants moderators in the loop rather than fully automated enforcement.

It is not sufficient on its own if the content hurting your community is contextual, multilingual, image-based, or written by someone deliberately spelling around your censor list. That is a description of what rule engines are for rather than a criticism. Running both is the normal answer.

For developers and platforms

Carl-bot is a Discord bot, so it does not help you moderate your own product. Supervisor has a REST API with SDKs for Python, JavaScript, Go, Rust and Java, plus a Platform API. The API reference has the endpoints.

The quickest way to judge any of this is on your own content. Paste a message that gets past a keyword filter into the live demo and see whether it gets caught, or add Supervisor to your server.