ModAI

ModAI is kesarAI’s agentic moderation assistant. A moderator describes what they want in plain language; the AI reads the relevant channel and member context, drafts a moderation plan— the exact sequence of actions it intends to take — and then waits. Nothing runs until a moderator clicks Confirm. It is a planner paired with an executor, not a chat bot: the planner proposes, a human approves, and only then does the executor touch the server.

It is the highest-privilege command in the bot. Behind the confirmation gate sits a registry of roughly 49 tools the executor can call — read tools that gather context (member lists, roles, audit history) and write tools that take action (ban, kick, mute, role changes, channel edits, and more). Because of that reach, ModAI is gated on an explicit mod_ai mod permission rather than the default Standard tier; administrators always pass, everyone else needs the grant.

Overview

ModAI exists for the requests that are awkward to express as a string of individual slash commands — “mute everyone who joined in the last hour and posted the same link,” “strip the booster role from anyone who stopped boosting,” “find the alt accounts.” You state the goal once; the planner figures out which members and which actions satisfy it.

The output is always a plan you can read before anything happens. Each step is labelled with its irreversibility, and any plan the planner judges destructive surfaces safety warnings and a second “really sure?” confirmation. The confirmation buttons are scoped to the moderator who invoked the plan — nobody else can click Confirm on your behalf.

How it works

The flow is a fixed pipeline:

1. Plan. The planner reads context with its read tools and drafts a structured plan: an intent summary plus an ordered list of steps, each tagged with the write tools it would call and how reversible it is.
2. Confirm. The plan is posted as an embed with Confirm and Cancel buttons. Nothing has touched the server yet. High-irreversibility plans require a double confirmation. The plan is also written to the audit log at this point, whether or not it is ever run, so there is a record of every proposal.
3. Execute. On confirm, the executor runs the steps in order through the tool registry, then edits the embed with a per-step result (success, failure, duration).

Some requests are pure questions — “who has the admin role?” In that case the planner produces a zero-step plan and ModAI just replies with the answer; there is nothing to confirm. You can reply to any plan or answer to ask a follow-up question about it, and ModAI uses the data it already gathered to respond. If your follow-up contains an action verb (“now ban them”), it is routed back through the planner as a fresh confirm-gated plan that reuses the IDs from the previous round.

Triggering ModAI

There are two ways to invoke it, and they behave identically:

• The /modai slash command, with your instruction in the prompt field.
• The mai message trigger — type mai <instruction> at the start of a message, for example mai mute @spammer 1h for raiding. The literal token mai must start the message and be followed by a space; it will not fire inside words like email.

Both paths run the same planner, the same quota check, and the same confirmation gate. Sending mai with no instruction returns a short usage hint rather than charging quota.

Seeking members

ModAI has a dedicated two-turn flow for scanning the member list, triggered by mai seek or mai seeking (and a few aliases like scan members). The first turn is deterministic: ModAI posts a filter menu and charges no quota. You then reply with your criteria in natural language, and that second turn runs through the planner.

The scan can filter members on a wide range of axes, including:

• Roles— with or without specific roles, role combinations, top role, or no roles at all.
• Account age— account older or newer than a given duration.
• Server join age— joined within a window, or before a date.
• Activity and XP— active or dormant for a period, level above or below a bound, XP or message-count thresholds, streak, prestige.
• Badges— presence or absence of HypeSquad, Active Developer, Early Supporter, and similar.
• Profile and display name— custom or server avatar, display name containing a substring.
• Status— in voice, bots only or no bots, pending screening, boosters or non-boosters.

Results can be sorted (most active, newest account, highest level, longest dormant) and limited. For large result sets ModAI splits its reply across multiple messages so the full list survives Discord’s per-message length cap.

Its own quota

ModAI does not draw from the per-user /ai point budget described on the AI quota page. It runs on a separate pool: 30 points per user, per rolling 24-hour window. The window starts on the first charge and is per-user, not midnight-aligned.

Each plan costs between 1 and 3 points, scaled to how much work it represents:

• 1 point— a simple plan: at most one step, at most one research round, low irreversibility.
• 2 points— a moderate plan: two or three steps or research rounds, medium irreversibility, or a handful of targets.
• 3 points— a complex plan: four or more steps, ten or more targets, high irreversibility, or any plan carrying safety warnings.

Quota is only charged once a plan is successfully produced — a failed AI call costs nothing — and only bot owners bypass the budget. The first turn of a seek flow is free because it is a deterministic menu, not an AI call. The AI quota page notes this separation too.

ModAI vs Autopilot

ModAI and Autopilot are easy to confuse because both are AI-driven moderation, but they sit at opposite ends of the spectrum:

• Autopilotis always-on and real-time. It scans every qualifying message’s media and URLs automatically and acts when a confidence threshold is crossed — no human in the loop per scan.
• ModAI is on-demand and confirm-gated. A human invokes it, reviews the proposed plan, and explicitly approves before anything runs.

A lighter, related tool is /ai modadvice: it scans recent messages in a channel and returns an advisory report of issues and recommended actions, but it never takes any action itself. Reach for modadvice when you want a read on a channel, Autopilot when you want passive enforcement, and ModAI when you want to direct a specific moderation task in your own words.