Slack MCP Server

    

Slack MCP Server

Catch up on Slack without reading it.

Unreads, threads, and search — in your agent’s context, from the session you already have.

npx -y @jtalk22/slack-mcp --setup

Claude Code Claude Desktop Cursor Copilot Windsurf Gemini CLI Codex CLI any stdio MCP client

▶ It’s Monday, 9:07—watch what blew up overnight · interactive walkthrough · setup guide

How it works · Why session auth · Grid & credentials · Install · 21 tools · Workflows · Local vs hosted

It’s Monday, 9:07. Slack has already formed opinions.

You ask “what blew up overnight?” and the agent reads the workspace instead of you. It reconstructs the 2 AM P1 from #incidents—owner, resolution, and the runbook step that is still wrong. It finds the printer PIN that has been waiting in #facilities for five months. Then it closes the handled loops—replies, reactions, read-state changes—only where you approve.

This is not screenshot automation. The agent calls Slack through a real MCP tool surface and receives typed results it can search, summarize, export, or act on.

Built past the demo

The difficult part is not another chat tool. It is the operating layer underneath: browser-session extraction that names its failure stages, a credential lifecycle built for rotation, full-fidelity reads, guarded writes, and typed workflow output. The code is plain JavaScript in this repository—audit it before trusting it with a session.

Two ways into Slack

Slack already knows who you are. The official path is a Slack-managed remote integration governed by workspace policy—a strong fit for organization-sanctioned deployments, documented by Slack with integration settings under admin control. This project is the direct local path: session-based auth from the browser session already in Chrome, local stdio, any stdio MCP client, and no Slack app or admin request. Same Slack identity. Same underlying permissions. A radically shorter path from your workspace to your agent.

Grid, credentials, and caching

Enterprise Grid. Grid runs aggressive session-anomaly detection. Browser-session automation can trip it, which flags the session and kills it, regardless of which tool drives the traffic. Outbound calls are paced by default to stay under burst thresholds (SLACK_MCP_MIN_REQUEST_INTERVAL_MS, default 350; SLACK_MCP_MAX_CONCURRENCY, default 3). Pacing lowers that risk; it does not remove it. On Grid, use the hosted OAuth tier or Slack's official MCP instead.

Credential extraction. --setup reads the newest xoxc- token from Chrome's on-disk LevelDB, snapshots the cookie SQLite database, retrieves Chrome Safe Storage from the macOS Keychain, and runs PBKDF2 + AES-128-CBC decryption locally. It writes the token file, Keychain entries, and non-secret metadata. It transmits nothing — the server talks to Slack and nowhere else.

This is the same access pattern credential stealers use. Chrome App-Bound Encryption exists to make this class of read harder, and infostealer families (Lumma, Vidar, Meduza) bypass it to lift live sessions. The mechanism here is comparable. What differs is that you run it, on your own machine, against your own session, and nothing leaves the host. The source is plain JavaScript in this repository; audit it before handing it a live session.

User cache. One cache exists: user-name lookups, populated on demand, 500 entries maximum, one-hour TTL. No message content, no channel history, and no persistent copy of the workspace is stored.

Install

Node 22 or 24 recommended. Node 20 remains supported for the v4 line.

npx -y @jtalk22/slack-mcp --setup

Prefer a persistent CLI: npm install -g @jtalk22/slack-mcp then slack-mcp --setup.

Then:

1. Pick your client in the setup guide.
2. Register the generated stdio command.
3. Fully restart the client.
4. Ask the agent to run slack_health_check.
5. A workspace name in the response means the connection is live.

Use the same server command everywhere:

{
  "command": "npx",
  "args": ["-y", "@jtalk22/slack-mcp"]
}

On macOS, setup can extract from Chrome and persist the selected storage backend. On other platforms, provide SLACK_TOKEN and SLACK_COOKIE through the client's environment configuration. Docker, HTTP, and detailed client examples live in docs/SETUP.md and docs/DEPLOYMENT-MODES.md.

21 tools: read, act, automate

The local surface ships 21 tools today: 12 read-only Slack operations, 4 write-path tools that each carry an MCP destructive annotation so clients can gate workspace writes, 3 local workflow tools including the catch-up itself, and 2 hosted-intelligence stubs that return a structured upgrade payload without making a Slack call. Four read tools accept include_rich_message_fields: true to surface attachments, blocks, files, reactions, and metadata—complete inputs and response contracts live in docs/API.md.

Advertising fewer tools. A client pays for the tool schema on every turn that carries it. SLACK_MCP_TOOLS=essentials advertises six tools — unread, history, search, thread, user lookup, send — costing roughly 985 estimated tokens of schema per turn against about 3,600 for all 21. SLACK_MCP_TOOLS=read advertises the 12 read-only Slack operations listed below, near 1,690. --tools=slack_x,slack_y takes an explicit set. The default stays all 21. Filtering changes what is advertised, not what is callable. Reproduce the numbers with node scripts/measure-tool-schema.js (a ~4-chars-per-token estimate).

Typed workflows: Slack in, JSON out

Bind a workflow kind to channels, priority people, retention, and cadence. slack_catch_me_up then reads that scope locally and hands your agent the evidence: which threads went unanswered and for how long, what your priority people said or were pinned on, which conversations actually moved. It does the gathering; your agent writes the summary against the contract below.

There is no server-side model in that path, because there does not need to be one — the client calling this server is already a language model. Hosted adds what genuinely needs infrastructure: indexed semantic retrieval, and running the same catch-up on a schedule while your laptop is shut.

npx -y @jtalk22/slack-mcp --apply-template oncall-handoff --channels C012345,C067890

Where credentials live

Resolution is deterministic; first hit wins:

1. SLACK_TOKEN + SLACK_COOKIE
2. token file (chmod 600)
3. macOS Keychain
4. Chrome extraction on macOS

Session credentials commonly rotate after one or two weeks. When Slack returns invalid_auth, not_authed, token_expired, token_revoked, account_inactive, or HTTP 401, run npx -y @jtalk22/slack-mcp --setup to recover locally. On macOS, slack_refresh_tokens or --refresh-tokens refreshes without leaving the client; the optional LaunchAgent in docs/SETUP.md keeps long-idle installations healthy.

{
  "mcpServers": {
    "slack-work": {
      "command": "npx",
      "args": ["-y", "@jtalk22/slack-mcp"],
      "env": { "SLACK_MCP_PROFILE": "work" }
    },
    "slack-personal": {
      "command": "npx",
      "args": ["-y", "@jtalk22/slack-mcp"],
      "env": { "SLACK_MCP_PROFILE": "personal" }
    }
  }
}

Free local when you’re driving. Hosted when it must drive itself.

When local control is enough, stop here—everything above is MIT-licensed and runs on your machine. The local product is complete, not a crippled trial: hosted earns the upgrade through continuity, intelligence, and collaboration, not by holding ordinary Slack access hostage. Hosted exists for work that must survive a rotating browser session:

• permanent OAuth;
• indexed semantic search;
• scheduled catch-up and triage;
• contract-validated workflow briefs;
• shared profiles and managed workspace continuity.

Local mode never contacts us; it runs on your machine and talks only to Slack. Hosted never receives a browser cookie; it runs on permanent OAuth, for work that has to survive a rotating session — unattended schedules, Enterprise Grid. Everything above is MIT-licensed and complete on its own.

See live hosted pricing →

Security and provenance

• Credential files are owner-only; Keychain-only mode keeps plaintext credentials off disk.
• Configuration fails closed for unknown storage modes and invalid profiles.
• Writes are atomic and shared credential state is process-locked.
• The local web server binds to localhost; workspace write tools carry destructive annotations.
• Every release publishes from CI with npm provenance.

Provenance: don't take my word for it

npm audit signatures

A clean result verifies that the package signatures and attestations trace back through the published release chain. Inspect the package before handing it a live Slack session. Full policy: SECURITY.md.

Documentation

Setup · API · Architecture · Compatibility · Deployment modes · Recipes · Troubleshooting · Roadmap

Contributing

PRs are welcome. Read CONTRIBUTING.md and run node --check on touched JavaScript before submitting.

License

MIT — see LICENSE.

Disclaimer

Not affiliated with Slack Technologies, Inc. This server uses browser-session credentials. Review your workspace's acceptable-use policy before running it.

npx -y @jtalk22/slack-mcp --setup