Discord MCP

Repository: github.com/sandraschi/discord-mcp

FastMCP 3.1 Discord MCP server with sampling (server-side Ollama / OpenAI-compatible or client LLM), agentic workflow (discord_agentic_workflow), skills (SkillsDirectoryProvider + skill://…/SKILL.md), prompts, and a 2026 fleet-style dashboard (React + Vite + Tailwind, dark glass shell).

Ports

• Backend: 10756 — REST (/api/v1/…) + MCP streamable HTTP at /mcp
• Frontend: 10757

Setup

1. Create a bot at Discord Developer Portal. Copy the bot token.
2. Invite the bot to your server (OAuth2 URL Generator, scope: bot).
3. Copy .env.example to .env in the repo root (discord-mcp/.env) and set DISCORD_TOKEN=... (no quotes). The backend loads this file on startup (python-dotenv). Restart the server after changing .env. If the token is already set in the system environment or Cursor MCP env, that value wins (.env does not override existing vars).
4. For local agentic sampling (when the MCP host does not provide sampling): run Ollama and optionally set DISCORD_SAMPLING_BASE_URL / DISCORD_SAMPLING_MODEL. Set DISCORD_SAMPLING_USE_CLIENT_LLM=1 if you want the host LLM to sample and the server handler only as fallback.

Run

• Webapp (backend + dashboard): from repo root run .\start.ps1 or start.bat, or from webapp/: .\start.ps1.
• Backend only: uv run python -m discord_mcp.server --mode dual --port 10756
• Stdio (MCP only): uv run python -m discord_mcp.server --mode stdio

Cursor MCP

• This repo as workspace root: .cursor/mcp.json registers discord-mcp (stdio). Set DISCORD_TOKEN in that file’s env or rely on your shell environment. Reload MCP / restart Cursor after edits.
• Global / multi-root merge: copy the discord-mcp block from cursor-config-template.json into your user MCP JSON (Windows: %USERPROFILE%\.cursor\mcp.json). Adjust cwd to your clone path if you are not using the default D:/Dev/repos/discord-mcp.

MCP HTTP (remote / Inspector)

• Endpoint: http://localhost:10756/mcp (streamable HTTP, FastMCP 3.1)
• Discovery / manifest: GET /api/v1/meta
• Health: GET /api/v1/health (includes sampling status)

Tools

• discord(operation=…) — Portmanteau: list_guilds, list_channels, send_message, get_messages, get_guild_stats, create_channel, create_guild, create_invite, list_invites, revoke_invite, list_members, get_member, list_active_threads, rag_ingest, rag_query. create_guild requires user OAuth2 (bot 403). list_members / get_member require GUILD_MEMBERS intent.
• discord_help(category=…, topic=…) — Multi-level help.
• discord_agentic_workflow(goal, ctx) — High-level goal via ctx.sample + tools (SEP-1577). Returns a structured dict (success, message, recommendations or error fields).

Prompts (registered)

discord_quick_start, discord_diagnostics, discord_moderation_playbook, discord_rag_workflow, discord_invite_operations

Skills

Bundled under src/discord_mcp/skills/<name>/SKILL.md, exposed as MCP resources (skill://…). Listed in the dashboard under Skills and via GET /api/v1/skills.

Resources

• resource://discord-mcp/capabilities — Short capability summary for clients.

API (REST)

• GET /api/v1/health — Health, token_set, rate limits, sampling, mcp_http_path
• GET /api/v1/meta — Tools, prompts, resources, skills root, sampling
• GET /api/v1/skills — Bundled skill previews (dashboard)
• GET /api/v1/guilds — List guilds (proxy to tool)
• GET /api/v1/guilds/{guild_id}/channels — List channels
• (See OpenAPI at /docs for full list.)

Discord HTTP 429 (API rate limits)

Discord enforces per-route limits on the REST API. On HTTP 429, the server now waits for retry_after (from the response body or Retry-After header) and retries the same request up to 5 times before returning an error. If you still see 429 after that, slow down (e.g. fewer parallel message fetches, smaller RAG ingest batches) or wait a minute—especially if "global": true appears in Discord’s response (rare; indicates a broader limit).

Safety and rate limits (anti-spam)

Write operations are rate-limited so the bot is not usable as a spambot:

• Messages: max 10/min global, max 3/min per channel, min 5s between sends (configurable).
• Channels: max 5 created per minute.
• Invites: max 5 created per minute (DISCORD_RATE_LIMIT_INVITES_PER_MINUTE).
• Message length: capped at 2000 (Discord max); override with DISCORD_MAX_MESSAGE_LENGTH.

Env: DISCORD_RATE_LIMIT_MESSAGES_PER_MINUTE, DISCORD_RATE_LIMIT_MESSAGES_PER_CHANNEL_PER_MINUTE, DISCORD_RATE_LIMIT_CHANNELS_PER_MINUTE, DISCORD_MAX_MESSAGE_LENGTH, DISCORD_MIN_MESSAGE_INTERVAL_SECONDS. When a limit is hit, the tool returns success: false, rate_limited: true, and an explanatory error. Current config is in GET /api/v1/health under rate_limit.

Documentation

• docs/README.md — doc index
• docs/TECHNICAL.md — architecture, env, Discord 429, MCP /mcp
• CHANGELOG.md — release notes

Standards

• FastMCP 3.1: instructions, sampling handler, skills provider, strict validation, streamable HTTP mount.
• Webapp: start.ps1, ports 10756/10757, dark glass layout, top bar status, activity log, Tools / Skills / Apps pages.