Slack MCP Server: What It Does and Whether It's Worth It
Slack's official MCP server lets AI tools search, read and post in Slack. Here is how to set it up and where it stops.

The Slack MCP server is the official way to connect Claude, Cursor, Perplexity and other AI tools to your Slack workspace. It's hosted by Slack, signs in with OAuth, and lets the AI search, read and post on your behalf without you copying threads into a chat window.
Below: what the server can and can't do, how to connect it to each client, the community alternative, and the security checks to run before you approve it.
What is the Slack MCP server?
The Slack MCP server is a remote MCP server built and run by Slack that exposes your workspace to AI clients as a set of tools. It lives at https://mcp.slack.com/mcp and speaks JSON-RPC 2.0 over Streamable HTTP (Slack Developer Docs). MCP, the Model Context Protocol, is the open standard AI clients use to call outside apps; the best MCP servers for work guide covers the rest of the landscape.
Slack opened the server and its Real-time Search API to a limited group in October 2025 and made both generally available on 17 February 2026. According to Slack's launch post, RTS queries and MCP tool calls grew 25x during that limited release, and more than 50 partners, including Anthropic, Google, OpenAI and Perplexity, were building on it (Slack, 2026).
Slack also works in the other direction. Slackbot can act as an MCP client and call tools in apps like Linear, Figma or Atlassian Rovo from inside a Slack conversation (Slack Help Center). This post is about the server: AI tools reaching into Slack.
What can the Slack MCP server do?
It can read almost everything you can see in Slack and take the common write actions, split into read tools and write tools (Slack Developer Docs).
Read tools
- Search messages and files, filtered by date, user and content type
- Search users, channels and emoji
- Read channel history and threads, and get reactions
- Read files, canvases (exported as Markdown) and lists
- Fetch user info, list channel members and list your channels
Write tools
- Send and schedule messages, or draft and preview them inside the AI client
- Create a channel or conversation
- Add reactions
- Create or update canvases, lists and list records
- Upload files
In practice that covers prompts like "summarize what I missed in #product since Tuesday and draft a reply to the open question" or "find the thread where we agreed the launch date." The AI searches, reads the thread and hands you a draft.
What do you need before you connect it?
You need a client backed by a registered Slack app, OAuth sign-in, and in most workspaces an admin's approval.
- Approved apps only. Only apps published in Slack's directory (the Slack Marketplace) or internal apps built for your workspace can use the server; unlisted apps are blocked. Admins can approve and manage every MCP client, and MCP activity shows up in audit logs.
- OAuth with user tokens. The server uses confidential OAuth with a client ID and secret, and PKCE for desktop clients. It doesn't support SSE connections or dynamic client registration, so a client can't register itself on the fly.
- Rate limits. Each tool inherits the limit of the Web API method behind it, from Tier 2 (about 20 calls a minute) to Tier 4 (about 100 a minute). Search and sending have their own special limits.
- IP allowlists. If the Slack app behind your client has one, MCP requests must come from allowed addresses.
Slack's docs don't tie the server to a specific Slack plan; the gate is admin approval of the client app.
How do you connect Slack MCP to Claude, Cursor and ChatGPT?
Claude and Cursor have official setup paths in Slack's docs; ChatGPT connects through OpenAI's own Slack app instead.
Claude Desktop and claude.ai
- Open Customize in the sidebar and go to Connectors.
- Click +, find Slack and add it.
- Complete the OAuth flow for your workspace (Slack Developer Docs).
On Team and Enterprise plans, an owner may need to enable the connector for the organization first.
Claude Code
- Install Slack's plugin with
/plugin install slackinside a session, orclaude plugin install slackfrom the terminal. - The plugin configures
https://mcp.slack.com/mcpfor you when it loads. - Sign in to Slack through the OAuth prompt.
If you've logged in to Claude Code with a claude.ai account, a Slack connector you added on claude.ai also shows up in /mcp automatically (Claude Code docs).
Cursor
- Open Cursor Settings and select the MCP tab, or use the Add to Cursor button on Slack's page.
- Add the server with Slack's published client ID:
{
"mcpServers": {
"slack": {
"url": "https://mcp.slack.com/mcp",
"auth": { "CLIENT_ID": "3660753192626.8903469228982" }
}
}
}
- Save, click Connect and authenticate with your workspace (Slack Developer Docs).
ChatGPT
ChatGPT is on Slack's partner list, but OpenAI documents its own Slack app rather than a custom MCP setup. On Plus, Pro, Business, Enterprise and Edu, find Slack under Apps or Plugins, select Connect and authorize your workspace. It searches channels and DMs you can already read, connects to one workspace at a time, and only uses semantic search on Slack Business+ or Enterprise+ with Slack AI Search turned on (OpenAI Help Center). Because the Slack server only accepts approved apps, pasting mcp.slack.com/mcp into ChatGPT as a custom server isn't a documented path. The general pattern for other servers is in How Does ChatGPT Use MCP?
What about korotovsky/slack-mcp-server?
It's the most popular community Slack MCP server, and it's not made by Slack: the README says plainly "This is not an official Slack product" (GitHub). It's MIT-licensed, runs locally over stdio, SSE or HTTP, and offers tools for channel history, thread replies, message search, unread messages, user groups and saved items.
The difference is how it signs in. Besides user and bot tokens, it accepts your browser session tokens (xoxc and xoxd), which act as your logged-in Slack session and skip the admin approval step the official server enforces. Posting, reactions and mark-as-read are disabled by default and turned on through environment variables. It's useful for personal automation or workspaces where you can't get an app approved, but check your workspace policy first: a session token in a config file carries everything your account can reach.
Is the Slack MCP server safe?
It's as safe as the permissions you grant and the content the AI reads, and Slack content is written by many people. The main risks:
- Prompt injection. Anyone who can post in a channel the AI reads can plant instructions in a message. If that same session can send messages or call other servers, an injected instruction can move data out.
- Broad reach. User tokens see your DMs, private channels and files. The AI gets all of it, not only the channel you had in mind.
- Write actions. Sending, scheduling and canvas edits are live. Approve write calls one by one in your client.
- Mixed servers. Slack's own docs warn to be careful when you connect it alongside other MCP servers with different security characteristics.
Admin approval and audit logs help a workspace keep track, and the full threat model is in Is MCP Safe?
Is the Slack MCP server worth it?
Yes, if you act in Slack every day: searching, catching up and drafting replies inside your existing permissions beats copy-pasting threads into a chat. It's supported by Slack, respects admin controls and works in the clients most people already use.
Two costs come with it. Its tool definitions load into the model's context before you type, and several heavy servers together crowd out room for the actual work (the MCP context tax). And it's a search tool, not memory: the AI only looks in Slack when a prompt sends it there, and it searches a stream where decisions sit mid-thread between lunch orders. It answers "what got said in this channel." It doesn't hand the model a current picture of what you're working on before you ask. That's the line between a connector and a context layer, drawn in full in Connectors vs MCP vs a Context Layer.
Where does Unabyss fit next to Slack MCP?
Unabyss does the second job. Slack is one of its live integrations: you pick the channels it learns from, it turns their threads into personal context alongside Gmail, Notion, Linear and your other apps, and serves that context to Claude, ChatGPT, Cursor and other MCP clients from one connection. It doesn't post, react or edit anything in Slack. In an August 2026 Unabyss benchmark against a baseline with the same tools connected, adding Unabyss cut cost by 36%, time by 38% and follow-up prompts by 60% (What Structured AI Context Is Worth).
Many people run both: Slack's server for searching and replying, Unabyss so every AI tool starts out knowing what those threads decided. See how it works.