Best AI Agent Skills for Slack and Team Chat (2026)

"Connect my agent to Slack" describes two opposite things, and picking the wrong one costs an afternoon.

In one direction, your coding agent reaches into Slack: it reads a thread, searches history, posts a summary. In the other, Slack reaches into your agent: someone types @Claude fix this in a channel and a coding session starts somewhere else. Different tools, different install paths, different admin conversations.

Sort out which one you want first. Everything below is organized around that split.

Direction one: your agent reads and writes Slack

This is the common case. You are working in a terminal, the context you need is in a channel, and you would rather not alt-tab.

The official plugin, and the one to start with

Slack maintains its own plugin for Claude Code and Cursor. It bundles an MCP server with a set of on-demand skills, which is why it covers more ground than a bare MCP connection.

Install in Claude Code: /plugin install slack@claude-plugins-official

Install in Cursor: /add-plugin slack

It ships six skills covering Slack messaging, search across messages and files and channels, the Web API, the Slack CLI, app creation, and Block Kit layout building. Alongside those are five slash commands for the workflows teams repeat: channel summaries, topic searches, announcements, standups, and digests.

Authentication is OAuth on first use. Worth knowing before you start: workspace admin approval is required before the MCP integration can be authenticated, so if you work somewhere with locked-down app installs, that is a conversation, not a checkbox. MIT licensed, maintained by the Slack API team.

A data-quality note we owe you. Our own directory listed this project's repository as slackapi/slack-mcp-server, and that URL now returns a 404. The maintained repositories are slackapi/slack-skills-plugin and slack-mcp-plugin. Separately, the old reference Slack MCP server that shipped under the Model Context Protocol organization has been moved to an archived repository and is no longer maintained. If you find a Slack MCP tutorial written before 2026, check which of these three it is describing.

korotovsky's Slack MCP server, and a caveat that matters

slack-mcp-server from korotovsky is the best-known community alternative. MIT, around 1.8k stars, supports stdio, SSE, and HTTP transports, and it is explicit that it is not an official Slack product.

Its distinguishing feature is also the thing to think hardest about. It supports three auth modes: bot tokens, user OAuth tokens, and browser session tokens extracted from your logged-in Slack web client. That third mode is described as running "without requiring additional permissions or bot installations."

Read that plainly: it works because it borrows your browser session rather than asking your workspace for permission. If you are a solo user on your own workspace, fine. If you are on a company workspace, that is a route around the controls your admins configured, and it will read as an incident if it surfaces during a security review. The maintainer's own default posture is cautious, which is a point in the project's favor: message posting, reactions, and mark-as-read are all disabled by default.

Choose this one when the official plugin does not fit, and use OAuth tokens when you do.

slackdump, for the archive problem

slackdump is a Go tool for exporting Slack workspaces, not an agent skill, though recent versions do ship MCP server functionality. It earns a place here because a real recurring need is "let my agent read three years of a channel," and streaming that through an MCP connection one page at a time is slow and expensive. Export once, point the agent at files, work locally.

Roughly 2.7k stars. Two things to know before you reach for it. It is licensed under AGPL v3, not a permissive licence, and plenty of companies have a blanket policy against AGPL code touching internal systems. And its own description is that it saves private and public messages "without admin privileges," which puts it in the same category as the browser-token mode above: technically effective, organizationally a conversation you should have first.

Treat any full workspace export as a sensitive artifact, because that is exactly what it is.

Microsoft Teams and the rest

mcp-teams-server from InditexTech covers Microsoft Teams. It is a community project at around 390 stars, which is a very different maturity tier from the official Slack plugin, and it is worth trialling on a low-stakes channel first.

There is no comparable Anthropic-official or Microsoft-official Teams plugin at the time of writing. If your organization is on Teams, expect more assembly.

The odd one out

Anthropic's own slack-gif-creator ships in the official skills repository and does one small thing: it generates GIFs sized and formatted for Slack. It has nothing to do with reading your workspace. It is on this page because people find it when searching for Slack skills and then wonder why it does not connect to anything.

Direction two: Slack drives your agent

The other model inverts the flow. You stay in Slack, mention Claude, and a coding session gets created elsewhere.

Claude Code in Slack routes a coding request to a session on the web. Progress updates post back into the thread, and when the work finishes you get a summary with buttons to view the session or open a pull request. Context comes from the thread you mentioned it in, or from recent channel messages if you mentioned it directly in a channel.

Setup is admin-then-user. A workspace administrator installs the Claude app from the Slack App Marketplace; each person then connects their own Claude account from the App Home tab, and connects at least one GitHub repository to Claude Code on the web. Claude is not added to any channel automatically. Someone has to type /invite @Claude in each channel where it should work, which turns channel membership into a usable access control.

There is a plan-level distinction worth checking before you plan a rollout. Anthropic is retiring this version for Team and Enterprise workspaces in favor of Claude Tag, which runs @Claude as the organization's shared identity with admin-configured access rather than under each individual's account. On Pro and Max plans, the version described above remains the setup path.

Known limits as documented: GitHub repositories only, one pull request per session, sessions count against the individual user's rate limits, and it does not work in direct messages at all.

The security point nobody puts in the headline

Anthropic's own documentation states it directly: when @Claude is invoked in Slack, it is given access to the conversation context, and "Claude may follow directions from other messages in the context, so users should make sure to only use Claude in trusted Slack conversations."

That is prompt injection described in plain language by the vendor. A Slack channel is a place where other people write text. If an agent reads that channel as instructions, then anyone who can post in it can influence what the agent does. This is not a reason to avoid the integration. It is a reason to keep it out of channels with external guests, customer shared channels, and any channel where automated systems post content scraped from elsewhere.

The same logic applies in direction one. An agent that searches Slack history is reading text written by other people, some of whom may not be who you think they are.

Side by side

What people build with this

Four patterns come up repeatedly, and they are worth knowing because they set expectations about which direction you need.

Standup and digest generation. Read a channel over a date range, summarize decisions and blockers, post it back. The official plugin ships slash commands for exactly this. Direction one.

Incident context assembly. When something breaks, the useful history is scattered across an alerts channel, a thread, and someone's DM. An agent that can search across all of it turns fifteen minutes of scrolling into one query. Pairs well with the incident response skills.

Bug triage from where bugs get reported. Someone reports a bug in a channel, you mention the agent in the thread, and the conversation context becomes the bug report. Direction two, and the strongest argument for it.

Announcement and Block Kit drafting. Formatting a Slack message properly is fiddly. The Block Kit skill in the official plugin validates layouts before you post, which is a small thing that saves a specific and annoying kind of rework.

Choosing

• On Slack, want the supported path, can get admin approval. Official Slack plugin. Start here.
• On Slack, personal workspace, no admin to ask. korotovsky's server, ideally with a user OAuth token rather than browser session tokens.
• Need history in bulk rather than live queries. slackdump, then work against the export.
• On Microsoft Teams. mcp-teams-server, with realistic expectations about maturity.
• Want to kick off coding work from Slack rather than pull Slack into your terminal. Claude Code in Slack, or Claude Tag if you are on Team or Enterprise.

For the surrounding workflow rather than the chat layer, see our roundups for project managers and customer service reps, or the sibling guide on sharing agent skills with your team.

Before you install

Two habits worth keeping, both learned the expensive way.

Check what a repository is before you assume how it installs. Several projects filed under "Slack skills" across various directories, including ours, are MCP servers, standalone CLIs, or archived reference implementations. Look for a SKILL.md file, or an explicit plugin install command in the README, before you plan around it.

And check the token you are handing over. A Slack integration with broad scopes can read every channel you can read. That is the point, and it is also the risk. Grant the narrowest scope that does the job, keep credentials out of the repository, and prefer the auth mode your workspace administrators would recognize if you had to explain it to them.