Discord Tavern-Style LLM Chatbot

Archived project (2023). This was one of my first big bot projects and is no longer maintained. It targets the LLM tooling landscape as it existed in mid-2023 — pinned dependency versions and the LangChain APIs used here have since moved on. Left up as a snapshot for reference and for anyone who wants to fork it.

A Discord bot that uses local LLMs and character cards (Pygmalion / TavernAI style) for casual roleplay chat. Supports both KoboldAI and Oobabooga's text-generation-webui as backends, and can auto-update the bot's avatar and display name to match the selected character.

Features

• Character cards — drop a Tavern-style PNG into Cards/ or a JSON into Characters/ and pick it at startup
• Two LLM backends — KoboldAI or Oobabooga (text-generation-webui with --api)
• Per-channel memory — sliding-window conversation history scoped to each channel
• Image captions — pasted images are run through BLIP and described back to the LLM
• Stop sequences — configurable per-bot stop tokens to cut off run-on generations
• Listen-only mode — bot reads but stays silent; toggle per channel via /listen
• Ping mode — when on, the bot replies to every message; when off, only when addressed (/pingmode)

Setup

1. Clone this repo.
2. Copy sample.env to .env and fill in the values (see below).
3. Run setup.bat (Windows) or setup.sh (Linux/macOS) to create a virtualenv and install dependencies.
4. Run run.bat / run.sh.
5. Pick a character at the prompt.

Discord only allows bots to be renamed twice per hour, so the auto-rename can fail silently on rapid restarts.

.env variables

Variable	What it does
DISCORD_BOT_TOKEN	Bot token from the Discord Developer Portal.
ENDPOINT	URL of your local LLM API. For Oobabooga running locally with --api, that's http://127.0.0.1:5000/.
CHANNEL_ID	Comma-separated list of channel IDs the bot is allowed to talk in.
CHAT_HISTORY_LINE_LIMIT	How many recent lines to keep in per-channel memory.
STOP_SEQUENCES	Comma-separated stop tokens. Use \n\n for blank-line stops.
MAX_NEW_TOKENS	Generation cap. KoboldAI maxes out at 512.
ALWAYS_REPLY	T = reply to every message; anything else = only when addressed.

Recommended LLM backend

Oobabooga's text-generation-webui with --api enabled.

Slash commands

If commands don't appear after startup, run /sync once to force a refresh.

Command	What it does
/sync	Force-refresh the slash command tree.
/reload <cog>	Reload a single cog without restarting the bot.
/test	No-op sanity check.
/listen	Toggle listen-only mode for the current channel.
/pingmode	Toggle whether the bot replies to every message or only when addressed.

Tip: making the .env file on Windows

Enable file-name extensions so you can rename sample.env.txt → .env:

Windows 11

Windows 10

License

MIT — see LICENSE.