agentcord

Run and manage AI coding agent sessions on your machine through Discord. Supports Claude Code and OpenAI Codex.

Each session gets a Discord channel for chatting with the agent and a tmux session for direct terminal access. Sessions are organized by project — create multiple sessions in the same codebase, each with their own channel.

Quick Start

npm install -g agentcord
mkdir my-bot && cd my-bot
agentcord setup
agentcord

The setup wizard walks you through creating a Discord app, configuring the bot token, and adding it to your server.

Requirements

• Node.js 22.6+ (uses native TypeScript execution)
• tmux (for terminal session access)
• Claude Code installed on the machine (@anthropic-ai/claude-agent-sdk)
• OpenAI Codex SDK for Codex sessions (@openai/codex-sdk)

How It Works

Discord message → SDK query() → coding agent
                                      ↓
Discord embeds ← stream processing ← async iterator

Hybrid architecture: The agent SDK handles structured streaming for Discord interaction, while each session also gets a tmux session. You can tmux attach and run claude --resume <session-id> to take over the same conversation in your terminal.

Project-based organization:

Discord Server
  └── my-api (category)
  │    ├── #claude-fix-auth        ← session in ~/Dev/my-api
  │    ├── #claude-add-tests       ← another session, same project
  │    └── #project-logs
  └── frontend (category)
       ├── #claude-redesign        ← session in ~/Dev/frontend
       └── #project-logs

Discord Commands

Sessions

Command	Description
/claude new <name> [directory]	Create a session with a Discord channel + tmux session
/claude list	List active sessions grouped by project
/claude end	End the session in the current channel
/claude continue	Continue the conversation
/claude stop	Abort current generation
/claude attach	Show tmux attach command for terminal access
/claude model <model>	Change model for the session
/claude verbose	Toggle tool call/result visibility
/claude sync	Reconnect orphaned sessions (tmux + existing provider channels, including Codex)

Shell

Command	Description
/shell run <command>	Execute a command in the session directory
/shell processes	List running background processes
/shell kill <pid>	Kill a process

Agent Personas

Command	Description
/agent use <persona>	Switch persona (code-reviewer, architect, debugger, security, performance, devops)
/agent list	List available personas
/agent clear	Reset to default

Project Config

Command	Description
/project personality <prompt>	Set a custom system prompt for the project
/project personality-show	Show current personality
/project personality-clear	Remove personality
/project skill-add <name> <prompt>	Add a reusable prompt template ({input} placeholder)
/project skill-run <name> [input]	Execute a skill
/project skill-list	List skills
/project mcp-add <name> <command>	Register an MCP server (writes .mcp.json)
/project mcp-list	List MCP servers
/project info	Show project config summary

Features

• Real-time streaming — Agent responses stream into Discord with edit-in-place updates
• Typing indicator — Shows "Bot is typing..." while the agent is working
• Message interruption — Send a new message to automatically interrupt and redirect the agent
• Interactive prompts — Multi-choice questions render as Discord buttons
• Task board — Agent task lists display as visual embeds with status emojis
• Tool output control — Hidden by default, toggle with /claude verbose
• Per-project customization — System prompts, skills, and MCP servers scoped to projects
• Agent personas — Switch between specialized roles (code reviewer, architect, etc.)
• Session persistence — Sessions survive bot restarts
• Terminal access — tmux attach to any session for direct CLI use

Configuration

The setup wizard (agentcord setup) creates a .env file. You can also edit it directly:

# Required
DISCORD_TOKEN=your-bot-token
DISCORD_CLIENT_ID=your-client-id

# Optional
DISCORD_GUILD_ID=your-guild-id        # Instant command registration
ALLOWED_USERS=123456789,987654321      # Comma-separated user IDs
ALLOW_ALL_USERS=false                  # Or true to skip whitelist
ALLOWED_PATHS=/Users/me/Dev            # Restrict accessible directories
DEFAULT_DIRECTORY=/Users/me/Dev        # Default for new sessions
CODEX_SANDBOX_MODE=workspace-write     # read-only | workspace-write | danger-full-access
CODEX_APPROVAL_POLICY=on-request       # never | on-request | on-failure | untrusted
CODEX_NETWORK_ACCESS_ENABLED=true      # true | false

You can also override Codex policy per session when creating/resuming via:

• /session new ... sandbox-mode:<mode> approval-policy:<policy> network-access:<bool>
• /session resume ... sandbox-mode:<mode> approval-policy:<policy> network-access:<bool>

Development

git clone https://github.com/radu2lupu/agentcord.git
cd agentcord
npm install
cp .env.example .env   # fill in your values
npm run dev            # start with --watch

License

MIT