Catch up on Slack without reading it. Unreads, threads, search. Browser-session or hosted OAuth.
This MCP server provides access to Slack to help users catch up without reading it directly. Based on a browser-session or hosted OAuth flow, it supports retrieving unreads, threads, and results from search, using the Slack API.
🛠️ Key Features
Unreads retrieval
Threads access
Slack search
OAuth-based or browser-session authorization
🚀 Use Cases
Review unread Slack messages without opening Slack
Follow up on specific conversation threads
Find information via Slack search
⚡ Developer Benefits
Implements a Model Context Protocol (MCP) server
Works in MCP-compatible developer-tools and agentic-ai workflows
Categorized for use with llm and model-context-protocol environments
⚠️ Limitations
Intended for Slack access; scope is limited to unreads, threads, and search
Requires either a browser-session or hosted OAuth authorization to operate
Checked against Slack's published plans and pricing on 2026-10-07.
A free workspace keeps 90 days of message history, allows 10 app integrations, and gets only Slack's basic AI. The AI people actually want — AI search across the workspace, channel recaps, Slackbot acting as an agent, AI workflow generation — starts on Business+ at $15 per user per month on annual billing ($18 month to month). Pro, at $7.25, buys unlimited history and unlimited apps, not the advanced AI. So a ten-person free workspace that wants an AI it can ask about its own Slack is looking at $1,800 a year, and the cheaper upgrade does not get them there.
This package gives that workspace an agent over the same Slack for nothing, running on the AI subscription its people already pay for. Three things follow from how it works:
It installs no Slack app, so it uses none of your 10 app slots. If your workspace is already at the cap, this is the only way left to add an integration at all.
It needs no admin approval, because it is not an app anyone has to approve. It reads Slack through your own browser session, with exactly the access you already have and nothing more.
It can export what you can still see. On a free plan, history older than 90 days stops being reachable — Slack's own free export covers public channels only and is admin-only, so DMs and private channels are not in it. slack_get_full_conversation writes them out with their threads while they are still inside the window. It cannot retrieve anything already past the line; nothing can.
The browser-session route is not a workaround to apologise for. For a workspace Slack's own AI does not serve, it is the only route there is.
Two ways into Slack
Slack's official MCP server is an app your workspace admin approves and controls. Use it when your organisation wants a sanctioned integration. This server uses the browser session already in Chrome instead, so there is no app to install and no admin to ask, and it runs as a local command for any stdio MCP client. It acts as you, with exactly your access.
Side by side: the managed integration path vs. the local session path
Slack official MCP
Slack MCP Server — local
Starting point
A Slack-managed remote integration
The Slack session already in Chrome
Workspace control
Governed by workspace integration settings
No Slack app or admin request for the local path
Transport
Streamable HTTP
Local stdio
Client surface
Slack's supported partner integrations
Any stdio MCP client
Authentication
OAuth
Existing browser session
Credential lifetime
Managed OAuth
Rotating session with health checks and refresh
Product surface
Broad Slack-native capabilities
20 tools across read, act, and automate
Protocol
Slack-managed
MCP 2026-07-28 and every 2025 revision, from the same binary
Runtime
Slack-managed
MIT code on your machine
Is the local path against Slack's terms?
Treat browser-session automation as an acceptable-use decision for you and your workspace. The server acts as your signed-in Slack identity and cannot read a channel you cannot read or act as another user. It does not evade server-side retention, DLP, compliance exports, or audit controls.
"No admin request" means there is no Slack app installation to approve. It does not mean workspace activity disappears from Slack's systems. If your policy requires a sanctioned integration, use Slack's official MCP server.
Install
Node 22 or 24 recommended. Node 20+ is supported and CI-tested.
bash
npx -y @jtalk22/slack-mcp --setup
Prefer a persistent CLI: npm install -g @jtalk22/slack-mcp then slack-mcp --setup.
Then:
Copy the stdio command --setup prints when it finishes.
Paste it into your client, or run the claude mcp add line it prints.
Fully restart the client.
Ask the agent to run slack_health_check.
A workspace name in the response means the connection is live.
Per-client configuration keys are in the setup guide.
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.
Client configuration matrix
Client
Configuration surface
Status
Claude Code
claude mcp add or ~/.claude.json
Documented
Claude Desktop
Desktop MCP configuration
Verified
Cursor
.cursor/mcp.json
Documented
GitHub Copilot
.vscode/mcp.json
Documented
Windsurf
~/.codeium/windsurf/mcp_config.json
Documented
Gemini CLI
~/.gemini/settings.json
Documented
Codex CLI
codex mcp add or ~/.codex/config.toml
Documented
Other clients
Any stdio MCP configuration
Protocol-compatible
20 tools: read, act, automate
Read channels, DMs, threads and search; act with a reply, a reaction or a read-state change; run the catch-up and typed workflows locally. The surface is 20 tools today: 13 read-only operations, 4 write-path tools that each carry an MCP destructive annotation so clients can gate workspace writes, and 3 local workflow tools. One of the reads is slack_session_report, which reads nothing from Slack: it returns what this process has done, as counts. Every tool does its work here — reads Slack or local state — and none is a placeholder for something you would have to pay for. 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.
Speaks MCP 2026-07-28 and every 2025 revision from the same binary — era-negotiated over stdio, stateless per request over HTTP (no Mcp-Session-Id; GET/DELETE answer 405). The claim is a test, not a sentence: test/mcp-era.test.js drives the real SDK client at both eras against the real entry points.
Advertising fewer tools. Every tool the server advertises sends its schema to the model on every turn, before you ask anything. SLACK_MCP_TOOLS=essentials advertises six (unread, history, search, thread, user lookup, send); SLACK_MCP_TOOLS=read the 13 read-only operations; --tools=slack_x,slack_y an explicit set. The default is all 20. Filtering changes what is advertised, not what is callable.
The full tool inventory
13 read-only operations
Tool
Purpose
slack_health_check
Verify credentials and workspace identity
slack_token_status
Inspect credential age, health, cache, profile, and storage state
slack_refresh_tokens
Refresh local credentials from the browser session on macOS—reads Slack, writes only local state
slack_list_conversations
List channels and DMs
slack_conversations_history
Read channel or DM history with optional rich fields
slack_get_full_conversation
Export complete history and threads
slack_search_messages
Search across the workspace
slack_get_thread
Read all replies in a thread
slack_users_info
Resolve a user
slack_list_users
Page through large workspace directories
slack_users_search
Search users by name, display name, or email
slack_conversations_unreads
Prioritize conversations with unread messages
Act in the workspace — 4 write-path tools
Tool
Purpose
MCP safety
slack_send_message
Send to a channel or DM
destructive
slack_add_reaction
Add an emoji reaction
destructive
slack_remove_reaction
Remove an emoji reaction
destructive
slack_conversations_mark
Mark a conversation read
destructive
Automate locally — 3 workflow tools
Tool
Purpose
slack_workflow_save
Save a typed workflow profile to ~/.slack-mcp-workflows.json
slack_workflows
List saved workflow profiles
slack_catch_me_up
Read a profile's channels since its cadence window and return structured catch-up evidence
slack_refresh_tokens reads Slack and writes only local credential state.
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 needs a server: running the same catch-up on a schedule while your laptop is shut.
Six editable templates ship in the package: oncall-handoff, support-triage, exec-monday, sprint-tracker, customer-feedback, and incident-room.
Playbooks for your agent
Three playbooks ship with the server, so an agent learns how to use these tools with nothing installed in the client.
slack-catch-up walks the agent through finding or creating the right workflow profile, running slack_catch_me_up, and citing the message behind every line. On a narrowed tool profile it builds the same answer from unreads, history and threads.
slack-workflow-profiles covers choosing the kind, channels, priority people and cadence for a profile, or starting from a template.
slack-send-safely keeps writes to what you ask for, explains a held send, and points to slack_session_report for what the session did.
Clients get them through the MCP Skills extension (SEP-2640: skills/list, skills/get), as skill://<name>/SKILL.md resources, or as prompts of the same name. A playbook is offered only when its tools are, so a read-only server never offers the sending playbook and the essentials profile never offers profile setup.
Add your team's own with SLACK_MCP_SKILLS_DIR=/path/to/skills, a folder of playbook folders that each hold a SKILL.md. They are checked at startup and served beside the bundled ones. An invalid one, or one that reuses a bundled name, is skipped with a warning on stderr. SLACK_MCP_SKILLS=off turns playbooks off.
Security and provenance
Every message this server returns becomes text in a model's context, next to your own instructions. Slack is a shared bus: a channel can hold Slack Connect participants from another workspace, guests, and apps relaying content from outside Slack entirely — and the same toolset that reads also writes. That adjacency is the risk worth naming.
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.
Author labels on every message
Each message carries origin (self, internal, external, bot, unknown) and author_trusted, derived from fields Slack already returns, so it costs no extra API call. bot outranks workspace membership, because an app inside your workspace routinely relays words written outside it. A batch containing outside authors carries an untrusted_content envelope naming the count.
It fails closed: an author this server cannot positively place inside your workspace is reported untrusted. The exception is a conversation Slack reports as not shared with any other workspace — one of those structurally cannot hold an outside author, so a colleague with no team id on their message is called internal rather than unknown. When that lookup fails, the label stays unknown.
SLACK_MCP_PROVENANCE=off|label|strict. label is the default and only adds keys. A per-call provenance argument may tighten the mode and never loosen it, because a caller asking for fewer labels may be repeating something it read.
What the labels cannot do
A label is advice to the model reading it. It is unsigned, it can be stripped by anything downstream, and it raises the cost of an injected instruction without making one impossible.
strict mode additionally holds slack_send_message once the session has read outside-authored text. The hold is released by confirm_untrusted_context, which the caller sets — so it interrupts the send and puts it on the record, but it cannot prove a human approved it.
Two surfaces are still unlabelled, and knowing that is better than assuming otherwise: user display names, real names and titles from the user tools, and channel topics and purposes from slack_list_conversations, are attacker-settable free text that reaches the model without an origin stamp.
Read-only mode, for when a label is not enough
bash
npx -y @jtalk22/slack-mcp --read-only # or SLACK_MCP_READ_ONLY=1
The four write-path tools are withheld from tools/listand refused at dispatch, because a client can call a name it was never offered. The filter is applied after the tool profile resolves, so no profile or custom tool list widens it back. An absent tool needs no trust in the model that would have called it.
The receipt
slack_session_report returns what this process actually did: messages read, how many were outside-authored and by which origin, writes attempted, writes held. Counts only — no message text, no channel name, no user id. A label is a claim about one message; this is the claim about the session, and it is how you check afterwards whether outside text reached the model and whether anything tried to send on the back of it.
Verify the release
bash
npm audit signatures
A clean result means the package you installed was built and signed by this repository's release workflow. Read the code before giving it your Slack session. Full policy: SECURITY.md.
How it works
Five parts sit under the tools: extraction from Chrome, credential storage, reads, writes, and typed workflow output. All of it is plain JavaScript in this repository.
The five parts
1. Extraction from Chrome
--setup turns the Slack identity Chrome already holds into a local MCP server:
finds the newest xoxc- token in Chrome's on-disk LevelDB;
snapshots the cookie SQLite database with its WAL sidecars;
retrieves Chrome Safe Storage from the macOS Keychain;
proactive health checks and automatic macOS refresh;
last-known-good in-memory credentials when persistence is temporarily unavailable;
isolated profiles for work and personal Slack;
fail-closed handling for invalid storage or profile configuration.
3. Reads
Read DMs and channels, search the workspace, export complete histories with threads, inspect unread state, and resolve users. Opt into blocks, attachments, files, reactions, metadata, and bot/app markers when text alone is not the real message.
4. Writes
Send a reply, add or remove a reaction, and mark a conversation read. Every write tool carries MCP's destructive annotation, so compatible clients ask before running it.
5. Typed workflow output
Save workflow profiles for incident rooms, executive briefs, support inboxes, launch watches, and your own. Profiles are local JSON; the hosted tier can deliver them as scheduled briefs.
Credentials
The server checks these in order and uses the first it finds.
Source
Where it comes from
1
SLACK_TOKEN + SLACK_COOKIE
your client's environment configuration
2
Token file
written by --setup, chmod 600
3
macOS Keychain
written by --setup
4
Chrome extraction
macOS only, run on demand
Session credentials expire, and how long they last varies: one written on 2026-07-04 still worked on 2026-10-07. slack_token_status shows the age of yours. When Slack rejects it (invalid_auth, not_authed, token_expired, token_revoked, account_inactive or HTTP 401), run npx -y @jtalk22/slack-mcp --setup. On macOS, slack_refresh_tokens or --refresh-tokens refreshes without leaving your client, and the optional LaunchAgent in docs/SETUP.md keeps an idle install working.
Enterprise Grid, extraction, 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 Slack's official MCP server.
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.
Storage modes and multi-workspace profiles
Mode
Behavior
auto
Token file plus Keychain backup
keychain-only
Keychain only; verified writes and no plaintext credential file
file
Owner-only token file; Keychain is never touched
The selected backend is remembered in non-secret metadata and used by the server, CLI, and optional refresh job. An unrecognized mode fails at startup instead of silently downgrading storage.
Each profile gets its own token file, Keychain entries, metadata, and lock. Add SLACK_MCP_CHROME_PROFILE when the workspaces live in different Chrome profiles.
Local and hosted
The package runs on your machine, talks only to Slack, and is free under the MIT license. The hosted tier runs on a schedule instead, so the brief arrives while your laptop is closed:
Not affiliated with Slack Technologies, Inc. This server uses browser-session credentials. Review your workspace's acceptable-use policy before running it.
Your Slack. Your agent. One command.
If this removes a Slack tab from your day, star the repository. Stars are how the next admin-blocked developer finds the local path.