Chloe-bot

A free, locally hosted, Perchance-backed Discord bot that runs entirely in your browser. A Tampermonkey userscript is the trusted bridge (it holds the bot token and talks to Discord's REST API); a Perchance generator page is the brain and control panel (AI text + image generation, persona, settings). No server, no hosting bill — if the tab is open, Chloe is alive.

Since v0.24.0 Chloe is multi-tab: the tab you set up is the queen (it owns the token, the Discord traffic, and all state), and you can spawn disposable worker tabs whose only job is lending their AI/image generators — so text replies and image generation run in parallel instead of queuing on one tab.

Requirements

• Chrome (or any browser with a userscript manager)
• Tampermonkey (or a compatible userscript manager) with the chloe-bridge.user.js script installed
• The Chloe control generator page open on perchance.org

Discord setup

1. Create an application + bot at the Discord developer portal.
2. Enable "Message Content Intent" under Bot → Privileged Gateway Intents. This is required — without it, Discord returns empty message content and Chloe will see people talking but never hear what they say.
3. Invite the bot to your server with at least: View Channel, Send Messages, Read Message History. Optional extras: Manage Messages (lets her pin the transparency notice), Ban Members (enables the human-confirmed permaban command), Attach Files (image replies).

Bot setup

1. Install the userscript and open the Chloe control generator. The Tampermonkey menu on that page has "Set bot token" — the token is stored only in the userscript's storage and never crosses into the page.
2. In Discord, enable Developer Mode (Settings → Advanced), right-click the channel you want Chloe in, Copy Channel ID, paste it into the panel, and click Set Channel.
3. Tip: on a first run in an existing channel, click Backfill history so she starts with context instead of amnesia.
4. Set any options (greetings, volunteering, images, auto-mod, beats…), then click Start.

Worker tabs (optional, new)

In the panel's Mods card, Spawn worker tab opens a second copy of the generator marked as a worker. Workers register with the queen within ~15 seconds and from then on AI and image jobs are scheduled onto whichever worker is idle (falling back to the queen's own tab automatically — including mid-job, if you close a worker while it's generating). Workers are disposable: close them whenever, shut them down from the panel, spawn more when the channel is busy. Only the queen talks to Discord; if the spawn does nothing, allow pop-ups for perchance.org.

Failover: if the queen tab dies (closed by accident, crashed, discarded), the worker tabs notice within a couple of minutes, elect exactly one of themselves as the new queen, and — if Chloe was running — resume her automatically. Closing your laptop lid is fine too: a machine waking from sleep is detected and does not trigger a false election. So with one worker tab open, "keep this exact tab alive" relaxes to "keep any one tab alive".

Multiple channels

The Setup tab takes extra channel IDs (comma-separated). Each channel gets its own independent Chloe: separate memory of people, separate moderation state and strikes, separate engagement mode (!chloe lockdown in one channel doesn't lock the others), separate persona note — all served by one bot token and one worker pool. A selector appears in the command bar to scope the roster, mod log, and mode controls to a channel.

The control panel

The panel is a small console: a sticky command bar (Chloe's heartbeat pulses while she runs) with Start/Stop and the channel selector, five tabs — Setup, Behavior, Moderation, People, System — and the live log always docked below.

Talking to Chloe

Mention her or use her name — both work:

• @chloe-bot how's it going?
• chloe, draw me a fox in a raincoat (image replies, delivered as native attachments). With image memory on, she remembers what she drew and understands natural follow-ups like "make it bigger", "another one", or "now at night".
• !chloe forget me — anyone can ask her to prune their own conversation history and stop being remembered (moderation records are kept); !chloe remember me re-enables memory later, without resurrecting anything from before.
• !chloe time / !chloe date — instant, no-AI answers from the operator's device clock (or, with the panel closed, an approximate reading from the latest message's timestamp).
• !chloe lang fr — set your own reply language; she reads and replies to you in it while her thinking and moderation stay in English. !chloe lang off returns to English.
• !chloe goal <what you're working on> — she keeps it as a lasting goal that follows you across channels and can resurface in a check-in.

See full-list-of-commands.md for everything.

She also volunteers into conversations she can help with (configurable), greets people proportionally to how well she knows them, and can post scheduled in-character "beats" when the room is active.

Personality: the panel has six dials (kindness, sarcasm, curiosity, playfulness, formality, verbosity) that shape her tone — 50 is neutral and silent. Mods can also anchor a style note straight from Discord: react 📌 to any message (e.g. "Chloe should be more playful today") and, within a few polls, that message becomes her current style guidance for the channel. The newest mod-anchored message wins; !chloe persona shows it and !chloe persona clear removes it. Notes are sanitized, length-capped, and treated strictly as tone guidance — they can't change her rules or moderation.

Moderator commands

Mods (managed from the panel) can use !chloe <verb> in-channel — or the !c short prefix, or emoji aliases:

Verb	What it does
ignore / unignore @u	She stops/resumes engaging with someone
timeout @u 1h [reason]	Temporary ignore with auto-expiry
softban / unsoftban @u	Persistent ignore
warn @u [reason]	Adds a strike (see below); warns @u reports the count
block @u [reason] / unblock @u	Permanently forget a user: blocked users are never scanned or remembered again, even if they keep talking. Unblock lets memory form again.
clear @u	Clean slate — state and strikes reset
note @u <text>	Attach a mod note
recap / status / help	Channel summary, engine status, command list (moderator-only in the current build)
lockdown / unlock / open	Mods-only mode / normal / reply-to-everyone

Permaban (irreversible: Discord ban + verified memory purge) is not a chat command — it can only be run from the control panel, with explicit confirmation. Auto-moderation never escalates to it.

Auto-moderation and the strike ladder

Optional rule list (panel-editable): each rule matches by text, regex, confusables (catches homoglyph evasion like "fr\u0435\u0435 nitro"), or link (matches inside URLs only), and applies a reversible action — ignore, timeout, softban, or warn. A warn adds a strike: strikes walk an escalating ladder (default: ignore → 10-minute timeout → 1-hour timeout → soft-ban) and decay with good behavior (default: one strike forgiven per 24h). Auto-moderation never escalates to anything irreversible, mods are exempt, and every action (manual or automatic) is logged with the target's recent lines for audit.

Output gates

Chloe is gated on what she's allowed to put in a message she sends, with five independent toggles in the Moderation tab (and safe defaults):

• custom emoji — <:name:id> emoji in her text (default: allowed)
• pings (@user / @role) — whether her messages resolve real pings (default: blocked)
• @everyone / @here — mass pings, a separate toggle (default: blocked)
• links (URLs) — http(s) links, which also prevents link auto-embeds (default: blocked)
• channel links (#channel) — <#id> channel references (default: blocked)

Pings use Discord's own allowed_mentions system — the platform itself decides what becomes a real notification, so even if a ping slips into her text it won't fire unless the gate is on. Emoji, links, and channel-links are scrubbed from her message content before it posts. Together these mean a jailbroken or prompt-injected Chloe still can't mass-ping the server, spam links, or @ channels unless a mod has explicitly allowed it.

DMs

Chloe can hold a two-way DM conversation. Because Discord's REST API gives no per-message "this is a DM" signal, you declare DM channels explicitly: the Setup tab has two boxes — Public channels and Private DMs — and any channel id you put in the DMs box is polled and treated as a one-to-one DM (she replies to every message there, no @-mention needed), with all the same moderation. She can also open a DM herself (dm.open from the panel). Turn replies on with the DM replies toggle.

What she knows about you in public follows you into a DM (facts, insights, familiarity) so a DM isn't a cold start — but the flow is one-way and DMs are isolated: anything learned in a DM (memories, goals, your forget/remember state) stays in that DM and never leaks back into public channels or into anyone else's DM. One Discord limit to know: she can't receive a cold DM from someone she's never messaged, because Discord only delivers those through the realtime Gateway, which this REST-polling bridge doesn't use — so DMs work for any conversation Chloe is part of, just not unsolicited inbound ones.

Personalities

When a mod anchors a persona note (the 📌 reaction, or the Personality card) that names a character — e.g. "Name: Seraphina, regal and aloof" or "you are Marcus the merchant" — Chloe becomes that character: she answers to that name, speaks in first person as them, and won't narrate that she's Chloe playing a role. Clear the note and she's Chloe again.

Grounding (what she actually knows)

A few optional features feed her true facts she would otherwise have to guess at, kept strictly as grounding (never as instructions that could change her rules):

• Date & time — with this on she's told the current date/time from your device clock (timezone name included; no IP, GPS, or address). !chloe time / !chloe date answer instantly with no AI. With the panel closed she falls back to the timestamp of the latest Discord message rather than guessing.
• Her own basics — her name, the command prefix people use, that she can be @-mentioned, and the summon reaction, so "how do I use you?" / "are you a bot?" get straight answers.
• An operator note — the "Tell Chloe a fact about right now" card lets you inject a short, true situational fact ("#support is for bug reports", "the server is in beta this week", "event Saturday"), with an optional expiry. She states it when relevant. Don't put anyone's location or private details there.

Privacy and transparency

• forget me lets anyone self-prune and stop being remembered; remember me re-enables memory without resurrecting anything from before.
• The panel can post and pin an editable transparency notice in the channel.
• The mod log survives purges; user memory does not survive forget me, block, or permaban.
• All memory lives in the operator's own browser (userscript + Perchance page storage). There is no server and nothing is sent to the author. See PrivacyPolicy.md.

Troubleshooting

• She replies to nothing / sees empty messages → Message Content Intent is off (step 2 of Discord setup).
• Spawn worker does nothing → pop-up blocker; allow pop-ups for perchance.org.
• She goes quiet when the computer sleeps → the queen tab must be open and the machine awake; she resumes on wake, and rate limits (429s) are honored automatically.

---

Made by west-ninja · therealwestninja