Slack Agent

• Slack Agent is an LLM-powered agent built using the LangGraph ReAct Agent workflow and MCP Server.
• Protocol Support: Compatible with A2A protocol for integration with external user clients.
• Secure by Design: Enforces Slack API token-based RBAC and supports secondary external authentication for strong access control.
• MCP Server: Uses the OSS korotovsky/slack-mcp-server — a Go-based Slack MCP server supporting multiple auth modes, transports (stdio/SSE/HTTP), and stealth mode.
• MCP Tools: Uses langchain-mcp-adapters to glue the tools from the Slack MCP server to the LangGraph ReAct Agent Graph.

Architecture

Detailed Sequence Diagram with Agentgateway

System Diagram

Sequence Diagram

---

Local Development Setup

Use this setup to test the agent against Slack.

Get Slack API Credentials

1. Go to Slack API
2. Create a new Slack app
3. Configure Bot Token Scopes:
   • channels:read (required — channel listing and cache)
   • channels:history (required — read channel messages)
   • users:read (required — user search and cache)
   • groups:history (private channel history)
   • im:history (direct message history)
   • mpim:history (group DM history)
   • search:read (message search)
   • chat:write (only if enabling message posting)
   • reactions:write (only if enabling reactions)
   • app_mentions:read (optional — for event subscriptions)
4. Enable Socket Mode and generate an app-level token
5. Install the app to your workspace
6. Save the Bot User OAuth Token

Add to your .env:

SLACK_BOT_TOKEN=<your-bot-token>

Local Development

# Navigate to the Slack agent directory
cd ai_platform_engineering/agents/slack

# Run the A2A agent (requires the MCP server to be running separately)
make run-a2a

To run the MCP server locally via Docker:

docker run -p 3001:3001 \
  -e SLACK_MCP_HOST=0.0.0.0 \
  -e SLACK_MCP_PORT=3001 \
  -e SLACK_MCP_XOXB_TOKEN=$SLACK_BOT_TOKEN \
  -e SLACK_MCP_ADD_MESSAGE_TOOL=true \
  ghcr.io/korotovsky/slack-mcp-server:v1.2.3 --transport http

MCP Server

The Slack agent uses the OSS korotovsky/slack-mcp-server which provides:

Read-Only Tools (enabled by default)

• conversations_history — Fetch channel messages with pagination by date/count
• conversations_replies — Retrieve thread messages
• conversations_unreads — Get all unread messages across channels
• channels_list — List workspace channels
• users_search — Find users by name/email
• conversations_search_messages — Search messages with multiple filters
• usergroups_list — List user groups/subteams

Write Tools (opt-in via env vars)

• conversations_add_message — Post messages (requires SLACK_MCP_ADD_MESSAGE_TOOL=true)
• reactions_add / reactions_remove — Manage emoji reactions
• conversations_mark — Mark channels as read (requires SLACK_MCP_MARK_TOOL=true)
• usergroups_create / usergroups_update / usergroups_users_update — Manage groups

Authentication Modes

• Bot token: Set SLACK_MCP_XOXB_TOKEN (maps from SLACK_BOT_TOKEN in docker-compose)
• User token: Set SLACK_MCP_XOXP_TOKEN
• Browser tokens: Set SLACK_MCP_XOXC_TOKEN + SLACK_MCP_XOXD_TOKEN (stealth mode, no bot install required)

Environment Variables

Variable	Description	Default
SLACK_MCP_XOXB_TOKEN	Slack bot token	(required, one of xoxb/xoxp/xoxc+xoxd)
SLACK_MCP_HOST	Listen host	127.0.0.1
SLACK_MCP_PORT	Listen port	13080
SLACK_MCP_ADD_MESSAGE_TOOL	Enable message posting	false
SLACK_MCP_MARK_TOOL	Enable mark-as-read	false
SLACK_MCP_LOG_LEVEL	Log level	info
SLACK_MCP_API_KEY	Bearer token for SSE/HTTP auth	—
SLACK_MCP_ENABLED_TOOLS	Comma-separated tool whitelist	(all read tools)

Safety Guardrails

The korotovsky/slack-mcp-server has built-in safety controls:

1. Write tools disabled by default — Message posting, reactions, and group management require explicit opt-in via environment variables.

2. Channel restrictions — SLACK_MCP_ADD_MESSAGE_TOOL supports three modes:

   • true — allow posting to all channels
   • C123,C456 — whitelist: only allow posting to specific channel IDs
   • !C789,!C012 — blacklist: allow all channels except these

3. Tool whitelisting — SLACK_MCP_ENABLED_TOOLS restricts which tools are registered (e.g. conversations_history,channels_list to allow only read operations).

Note: There is no built-in regex filtering or rate limiting. For additional protection, rely on Slack API token scopes (e.g., remove chat:write scope to prevent posting entirely) and the channel restriction features above.

Example Use Cases

Ask the agent natural language questions like:

• Channel Operations: "List all channels in the workspace"
• Message Management: "Send a message to #general saying 'Hello team!'"
• User Management: "Show me information about user 'john.doe'"
• Reaction Management: "Add a thumbs up reaction to the latest message in #announcements"
• Thread Management: "Reply to the message in #bug-reports about the login issue"
• Search Operations: "Find all messages containing 'deployment' in the last week"