slack-desktop-mcp

Runs korotovsky/slack-mcp-server against the session your Slack desktop app is already signed into. No Slack API tokens to create, store, or rotate.

Prerequisite: bunx (Bun). The MCP server binary is downloaded on first use.

Register it

SLACK_WORKSPACE_MATCH selects the workspace when you're signed into several; omit it and the last-active one wins.

Claude Code

claude mcp add slack --env SLACK_WORKSPACE_MATCH=acme -- bunx slack-desktop-mcp

Claude Desktop — edit its config, then fully quit and relaunch it (the file is read only at startup):

| OS | Path | | --- | --- | | macOS | ~/Library/Application Support/Claude/claude_desktop_config.json | | Windows | %APPDATA%\Claude\claude_desktop_config.json | | Linux | ~/.config/Claude/claude_desktop_config.json |

{
  "mcpServers": {
    "slack": {
      "command": "bunx",
      "args": ["slack-desktop-mcp"],
      "env": { "SLACK_WORKSPACE_MATCH": "acme" }
    }
  }
}

Spawn/ENOENT error means the GUI-launched app didn't inherit a PATH containing bunx; use its absolute path (which bunx) as command.

slack-desktop-mcp --help lists the flags.

How it works

1. A cached session that still passes auth.test is used as-is — Slack is never touched.
2. Otherwise Slack is launched with --remote-debugging-pipe and CDP over the inherited fds (no TCP port) reads the d cookie and the xoxc token out of localStorage, then the launched instance is closed again. Both reads are polled: the page target exists before the client has written its session.
3. slack-mcp-server is fetched from GitHub Releases, cached per version, and exec'd with the tokens in its environment — it becomes the MCP server on this process's stdio, which is why all logging here goes to stderr.

Slack already running elsewhere owns its single-instance lock, so step 2 fails fast rather than disturbing that window; quit Slack and retry.

It relaunches itself once under bun --use-system-ca, so step 1 trusts the machine's certificate store; a TLS-intercepting proxy fails it otherwise. If step 1 can't complete at all, the cached session is used unverified rather than treated as expired — relaunching Slack can't fix an unreachable slack.com.

Step 2 costs a Slack launch plus its sign-in, which can exceed an MCP client's connect timeout; Claude Code takes MCP_TIMEOUT (ms) if a cold start trips it.

Failures also raise a native dialog, because MCP clients don't surface a stdio server's stderr. --no-ui keeps it headless.

~/.cache/slack-desktop-mcp/ holds session.json (mode 600), the downloaded binaries, and refresh.lock. Only one process may drive Slack at a time: the holder clears the cached session on acquire and keeps the lock while its dialog is up, so the cache only ever holds a session some refresh completed. Anyone who blocks on the lock never launches Slack itself — it re-reads the cache when the lock frees, and exits nonzero on stderr alone if the holder produced nothing. A lock older than 120s is treated as abandoned.