Discord Community Moderation Bot

A lean, experimental Discord bot that reads messages in channels it can access, classifies each one against your community rules using the BlueClaw (OpenAI-compatible) API, and replies to violations with an embed quoting the message and a fixed flag notice.

When flagged, the bot replies with an embed:

<the offending message, quoted in full; truncated only if the embed would exceed Discord's limit>

This message does not align with our community rules. If you think this was flagged mistakenly, please submit a PR here

(community rules → RULES_URL, PR here → PR_URL. Masked links are used because they render reliably inside embeds, unlike plain message content.)

Stack

• Python 3.11+
• discord.py — gateway + message events
• httpx — BlueClaw API calls
• python-dotenv — .env loading

Setup

python -m venv .venv && source .venv/bin/activate
pip install -r requirements.txt
cp .env.example .env   # then fill in the values

In the Discord Developer Portal, enable the Message Content Intent (privileged) and invite the bot with Read Messages + Send Messages permissions.

BlueClaw API key

The classifier runs on BlueClaw, which exposes an OpenAI-compatible endpoint. To try the bot, sign up at blueclaw.network, create an API key from your account dashboard, and put it in .env as BLUECLAW_API_KEY (with BLUECLAW_BASE_URL pointing at the endpoint shown there). The key is kept out of git and never logged.

Configuration (.env)

Var	Purpose
DISCORD_BOT_TOKEN	Bot token. Required to run live.
BLUECLAW_BASE_URL	https://openai.blueclaw.network/v1
BLUECLAW_MODEL	Qwen/Qwen3.6-35B-A3B-FP8
BLUECLAW_API_KEY	BlueClaw key. Kept out of git; never logged.
BLUECLAW_MAX_TOKENS	Completion budget (default 4096). Qwen3 spends tokens reasoning — too low truncates the reply to null before any JSON, especially with long rule sets.
RULES_URL	Link target for "community rules". Defaults to rules.md on GitHub.
PR_URL	Link target for "PR here". Defaults to the GitHub repo.
DRY_RUN	true logs the would-be flag instead of posting.
CHANNEL_IDS	Comma-separated allowlist. Empty = all visible channels.
MIN_MESSAGE_CHARS	Skip messages shorter than this.
CONTEXT_MESSAGE_COUNT	Previous human channel messages to include as context for each moderation check (default 4).
LOG_LEVEL	INFO / DEBUG / ...

Edit rules.md to change moderation behavior — its full text is injected into the classifier prompt at startup.

Run

python bot.py

Start with DRY_RUN=true to watch decisions in the logs before posting anything.

Behavior notes

• Fail-open: any API or JSON-parse error is treated as "not out of line", so the bot never blocks or spams on errors.
• Skips: its own messages, other bots, DMs, and empty/too-short messages.
• Conversation context: each moderation call includes up to CONTEXT_MESSAGE_COUNT previous human messages from the same channel, oldest-to-newest, plus the current message being judged.
• No structured-output dependency: the classifier is prompted for strict JSON and parsed defensively (direct / fenced / first-object); non-boolean verdicts are rejected.
• Qwen3 reasoning: the prompt prepends /no_think to suppress chain-of-thought and BLUECLAW_MAX_TOKENS (default 4096) leaves headroom so the JSON is never truncated.
• No secret logging: the API key is never written to logs.

Contributing moderation feedback

If the bot flags a message incorrectly, or misses a message that should have been flagged, please open a moderation rule feedback issue with:

• the message example
• the prior conversation context, if relevant
• what happened vs. what you expected
• evidence and reasoning tied to rules.md
• a proposed rule fix, ideally followed by a PR updating rules.md

See CONTRIBUTING.md for details.

Tests

pip install pytest
python -m pytest -q

Covers embed formatting/truncation and JSON decision parsing — both pure, no network.