Discord MCP

Connect your Discord bot to MCP clients -- list servers, send messages, moderate members, search message history with RAG, and run agentic workflows from Cursor or Claude Desktop.

v0.3.0 · 43 operations · FastMCP 3.2 · Comms lane · Releases

Contents

• Features

• Quick start

• What you can do

• Ports

• Documentation

• Requirements

Related MCP server: Discord MCP Server

Features

• 43 Discord operations in one portmanteau tool -- messaging, moderation, roles, webhooks, audit log, channel management (create/update/move, pins, threads), RAG

• Fleet web dashboard -- servers catalog with global selection, channel tree, decoded audit log, servers, channels, messages, agentic chat, LanceDB search (ports 10756 / 10757)

• Agentic workflow -- describe a goal; server uses sampling + tools (SEP-1577); model auto-resolves to an installed Ollama model

• Dual transport -- stdio for IDE hosts, streamable HTTP at /mcp for remote clients

• Built-in safety -- anti-spam rate limits, Discord 429 auto-retry, bind localhost only

• Bundled skills & prompts -- moderation playbook, RAG workflow, ops guides

Quick start

git clone https://github.com/sandraschi/discord-mcp
cd discord-mcp
.\start.ps1

1. Copy .env.example → .env and set DISCORD_TOKEN (bot setup)

2. Open dashboard http://127.0.0.1:10757 · API http://127.0.0.1:10756

Other install paths (just, Cursor, Claude Desktop, no-git): INSTALL.md

What you can do

List servers and post a message

List my Discord servers, then send "Fleet check-in OK" to channel #general in the first server.

Moderation assist

Show recent audit log entries for server 123456789 and summarize ban/kick events from the last 24 hours.

Search ingested history

Ingest the last 50 messages from #dev, then answer: what did we decide about the CI workflow?

Ports

Documentation

Fleet central mirror: mcp-central-docs/projects/discord-mcp

Requirements

• Windows (primary; fleet dev target) or macOS/Linux with Python 3.12+

• uv for Python deps · Node 20+ for the webapp

• Discord bot token -- free at Discord Developer Portal

• Optional: Ollama for local agentic sampling when the MCP host has no LLM

Server limit: Unverified bots can join 10 servers max. To raise the cap, verify your bot in Developer Portal → your application → Settings → Bot → Verification (requires description, screenshots, privacy policy). Once verified, the bot can join up to 100 servers.

How server owners add the bot: Send them the invite URL. They open it, select their server, choose which permissions to grant (or keep the defaults from the URL), and Authorise. The bot appears in their member list with a BOT tag.

What the bot can do depends on the permissions each server owner granted. The invite URL above requests Administrator (permissions=8), which gives full access. But the server owner can reduce permissions during authorisation. With full Admin the bot can read/search/post messages, moderate, manage channels/roles/webhooks, and view audit logs. With limited permissions it can only do what was allowed. The permissions are per-server, so the bot may be admin in one server and read-only in another.

AI & LLM Policy on Discord

Discord allows bots to use LLMs for moderation, search, summaries, and community tools. Using discord-mcp (reading accessible messages, RAG search, LLM-generated responses in your own servers) is within Discord's permitted use.

Allowed:

• Bots that use LLMs to process, summarize, search, or respond to messages

• Auto-moderation, content review, and audit-log analysis via LLMs

• RAG (vector search) over your own channel history

• Agentic workflows that call tools and generate replies

Not allowed (can get your bot suspended):

• Scraping user messages for external AI training data -- Discord prohibits crawling for model training

• Impersonating users with AI-generated content mimicking real people

• Spamming with LLM-generated flood messages (our rate limits prevent this)

• Selling access to Discord data through AI services

Bottom line: LLM as a tool to serve your community = fine. Discord as a free training data source = not fine. discord-mcp is designed for the former -- all operations are scoped to channels the bot can see, with rate limits that prevent abuse.

Rate Limits & Anti-Spam

The server enforces two layers of rate limiting to prevent abuse and Discord API bans.

Server-Side Anti-Spam (configurable)

These limits are checked before the request reaches Discord. If you hit one, the API returns a 429 with a message telling you which limit and how to override it.

Discord API Rate Limits (automatic retry)

Discord itself enforces per-route rate limits (HTTP 429 with retry_after). The server automatically retries up to 5 times with exponential backoff, respecting Discord's Retry-After header. If all retries are exhausted, a structured error is returned.

Fleet Crossconnects (Companions)

discord-mcp works completely standalone. You can optionally connect it with companion servers in the sandraschi fleet to unlock extended features:

Self-Contained Companions: Fleet companions operate independently. Installing companions does not trigger transitive dependency chains.

Repository: github.com/sandraschi/discord-mcp