llmcord

Talk to LLMs with your friends!

llmcord transforms Discord into a collaborative LLM frontend. It works with practically any LLM, remote or locally hosted.

Features

Reply-based conversations

Just @ the bot to start a conversation and reply to continue. Build conversations with reply chains!

The reply chain is the conversation history, stored entirely in Discord. No database required.

You can:

• Branch conversations endlessly
• Continue other people's conversations
• @ the bot while replying to ANY message to include it in the conversation

Additionally:

• When DMing the bot, conversations continue automatically (no reply required). To start a fresh conversation, just @ the bot. You can still reply to continue from anywhere.
• You can branch conversations into threads. Just create a thread from any message and @ the bot inside to continue.
• Back-to-back messages from the same user are automatically chained together. Just reply to the latest one and the bot will see all of them.

---

Model switching with /model

llmcord supports remote models from:

• OpenRouter
• OpenAI
• xAI
• Google

Or run local models with:

• LM Studio
• Ollama
• vLLM

...Or use any other OpenAI /v1/chat/completions compatible API server.

---

Per-user memory

Off by default. When enabled, a second model reads each exchange and records durable facts about the people talking to the bot, which are injected into the system prompt on later messages.

Two kinds of thing get stored:

• Self-memory: what you have told the bot about yourself.
• Attributed claims: statements you made about other people, stored against you as the source and against them as subjects. The bot never treats a claim as fact, it records who said it.

In a shared channel the bot always sees your own self-memory and your own claims. It sees other participants' memory only when memory_shared is true, because a user's memory can contain things they told the bot privately in DMs.

Use /memory to see everything stored about you and /forget to delete it, covering both what you told the bot about others and what others told the bot about you.

---

And more

• Supports image attachments when using a vision model (like gpt-5, grok-4, claude-4, etc.)
• Supports text file attachments (.txt, .py, .c, etc.)
• Customizable personality (aka system prompt)
• Distinguishes users via their Discord IDs
• Streamed responses (turns green when complete, automatically splits into separate messages when too long)
• Hot reloading config (you can change settings without restarting the bot)
• Displays helpful warnings when appropriate (like "⚠️ Only using last 25 messages" when the customizable message limit is exceeded)
• Caches message data in a size-managed (no memory leaks) and mutex-protected (no race conditions) global dictionary to maximize efficiency and minimize Discord API calls
• Fully asynchronous
• 1 Python file, ~300 lines of code

Instructions

1. Clone the repo:

   git clone https://github.com/jakobdylanc/llmcord
   cd llmcord

2. Set up config.yaml. Every available setting is listed under Configuration below.

3. Run the bot:

   No Docker:

   python -m pip install -U -r requirements.txt
   python llmcord.py

   With Docker:

   docker compose up

   Memory is kept in a named llmcord-data volume, so the same compose file works under Docker and rootless podman. If you are upgrading from a version that bind-mounted ./data, copy your existing data/memory.db into the volume before starting.

Configuration

Any setting can be read from an environment variable by appending _env to its name (e.g. bot_token_env: DISCORD_BOT_TOKEN).

Discord settings

Setting	Description
bot_token	Create a new Discord bot at discord.com/developers/applications and generate a token under the "Bot" tab. Also enable "MESSAGE CONTENT INTENT".
client_id	Found under the "OAuth2" tab of the Discord bot you just made.
status_message	Set a custom message that displays on the bot's Discord profile. Max 128 characters.
max_text	The maximum amount of text allowed in a single message, including text from file attachments. Default: 100,000
max_images	The maximum number of image attachments allowed in a single message. Default: 5 Only applicable when using a vision model.
max_messages	The maximum number of messages allowed in a reply chain. When exceeded, the oldest messages are dropped. Default: 25
use_plain_responses	When set to true the bot will use plaintext responses instead of embeds. Plaintext responses have a shorter character limit so the bot's messages may split more often. Default: false Also disables streamed responses and warning messages.
allow_dms	Set to false to disable direct message access. Default: true
permissions	Configure access permissions for users, roles and channels, each with a list of allowed_ids and blocked_ids. Control which users are admins with admin_ids. Admins can change the model with /model and DM the bot even if allow_dms is false. Leave allowed_ids empty to allow ALL in that category. Role and channel permissions do not affect DMs. You can use category IDs to control channel permissions in groups.

LLM settings

Setting	Description
providers	Add the LLM providers you want to use, each with a base_url and optional api_key entry. Popular providers (openrouter, openai, ollama, etc.) are already included. Only supports OpenAI /v1/chat/completions compatible APIs. Some providers may need extra_headers / extra_query / extra_body entries for extra HTTP data. See the included azure-openai provider for an example.
models	Add the models you want to use in <provider>/<model>: <parameters> format (examples are included). When you run /model these models will show up as autocomplete suggestions. Refer to each provider's documentation for supported parameters. The first model in your models list will be the default model at startup. Some vision models may need :vision added to the end of their name to enable image support.
system_prompt	Write anything you want to customize the bot's behavior! Leave blank for no system prompt. You can use the {date} and {time} tags in your system prompt to insert the current date and time, based on your host computer's time zone. It is recommended to include something like "User messages are prefixed with their Discord ID as <@ID>. Use this format to mention users." in your system prompt to help the bot understand the user message format.

Memory settings

Setting	Description
memory_enabled	Set to true to store and inject per-user memory. Default: false
memory_db_path	Where the SQLite memory database lives. Default: data/memory.db In the container /app is mounted read-only, so this must stay inside data/.
memory_model	The <provider>/<model> used to extract memory from each exchange. Memory extraction stays off until this is set, even when memory_enabled is true.
memory_shared	When true, the bot also sees other participants' memory and the claims others recorded about them. Default: false Off by default because a user's memory can contain things they told the bot privately in DMs.
max_memory_text	The maximum size of a single injected memory block. Default: 2000
max_memory_total	The maximum size of all injected memory blocks combined. When exceeded, the least relevant blocks are dropped: other participants' memory first, then the oldest claims. Default: 6000
memory_extraction_max_messages	How many recent transcript messages are fed into each extraction pass. Default: 8
memory_max_claims_per_extraction	The maximum number of claims about other people that a single extraction pass can record. Default: 20
memory_extraction_prompt	Overrides the built-in prompt used to extract memory. Leave blank to use the default.

Notes

• If you're having issues, try my suggestions here

• PRs are welcome :)