AgntUX Slack

Turn your Slack workspace into an AgntUX knowledge store, and let Claude draft replies on demand — but only ever send them after you confirm.

What it does

• Reads every channel and DM you have access to, every 30 minutes during weekday work hours (7am–7pm, your local timezone). DMs, @mentions, and active threads are time-sensitive while you're working but don't need overnight or weekend polling — quiet hours conserve tokens.
• Extracts entities: Slack users (people), organisations referenced in shared links, workstreams from your # Glossary, and recurring topics.
• Triages action items: response-needed (DMs, @mentions), deadlines, decisions buried in long threads, knowledge updates from pinned messages and canvases.
• Handles threads correctly: every reply, every entity row, and every action item links back to the parent thread's (channel_id, thread_ts). New replies on old threads are caught via a tracked-threads registry — no missed activity.
• Drafts replies on demand. Click a Draft a reply button on an action item; the plugin fetches the thread, drafts a reply in chat, and shows an inline compose card. The actual slack_send_message call only fires after you confirm via the Send button.
• Summarises threads to Slack canvases. Click Summarise to canvas on an action item; the plugin renders an editable canvas card with TL;DR, decisions, open questions, and participants. The Create button triggers the canvas post back to the thread.

Install

1. Make sure the Slack Connector is connected to your host (e.g., via the Anthropic Connectors marketplace at https://mcp.slack.com/mcp). agntux-slack does not authenticate with Slack itself — it talks to the host-installed Slack MCP server.
2. Install AgntUX Core if you haven't already.
3. Install AgntUX Slack from the marketplace.
4. Run /agntux onboard (or re-run it if your tenant is already onboarded). The flow handles agntux-slack's schema review automatically: personalization dispatches the data-architect's Mode B during the per-plugin interview, and the architect reads our schema proposal directly from plugins/agntux-slack/marketplace/listing.yaml → proposed_schema, walks you through it in plain language, and writes the approved contract at <agntux project root>/data/schema/contracts/agntux-slack.md. Zero manual schema-review steps.
5. Onboarding's State A wrap-up auto-fires /agntux sync agntux-slack for the first synchronous bootstrap; the ongoing schedule (every 30 min during weekday work hours) takes over after.
6. To trigger a sync manually any time, run /agntux-slack (or /agntux-slack sync, or /agntux sync agntux-slack from the core namespace). Or ask a live question: /agntux-slack what's been happening in #implementations. To revisit the architect's contract decisions later, run /agntux schema review agntux-slack (only needed if you want to change the approved contract — it is NOT a required install step).

Configuration

Bootstrap window: on the first run the plugin ingests Slack messages from the last 7 days — overrides the P3 §6.1 default of 30 because Slack volume is much higher than notes/email. To override, add bootstrap_window_days: N to the frontmatter of <agntux project root>/user.md (range 1–365).

Channel coverage: the plugin discovers channels three ways on every run — channels you've authored in, channels you've been @mentioned in, and DMs/MPIMs. Once a channel surfaces, it stays in the per-channel cursor map and is polled on every run.

First-run consent: the discovery sweep uses slack_search_public_and_private, which the host requires you to approve on first call. If you see a kind: auth entry in data/learnings/agntux-slack/sync.md → errors, grant the connector's search permission in your host and re-run /agntux-slack (or /agntux-slack sync).

Triage preferences: edit <agntux project root>/user.md → # Preferences to control which Slack messages generate action items. Add patterns to ## Always action-worthy or ## Usually noise. To allow specific bot messages (the default skips all bot traffic), add a per-plugin instruction in <agntux project root>/data/instructions/agntux-slack.md under # Always raise — e.g., bot_id:B01ABC to allow a GitHub PR bot.

Suggested-action flow

Action items raised by the sync sub-command (skills/agntux-slack/SKILL.md

• reference/sync.md) ship three buttons by default (Draft a reply, Schedule a reply, Open in Slack) plus an optional fourth (Summarise to canvas) for thread-summary-worthy items. Snooze, "Stop raising items like this", and "Mark done — already handled in Slack" are intentionally NOT plugin-authored — all three are redundant with built-in agntux-core triage chrome (Snooze button with 24h preset; Stop-raising in the Details modal; primary Done button on every card; Dismiss → "Completed externally" outcome). Plugin-side duplicates were retired in 3.0.0 for Snooze / Stop-raising and in 4.0.0 for "Mark done".

When you click Draft a reply, the host routes a ux: Use the agntux-slack plugin to open the reply composer for action {id}. prompt back to this plugin. The host's description-based auto-routing matches the prompt against the compose_view MCP tool directly — no skill round-trip — and the iframe renders inline with the pre-composed draft already loaded.

1. The sync skill runs at scheduled cadence and pre-composes the draft body for every Draft a reply-bearing action item, informed by user.md → # Preferences, data/instructions/agntux-slack.md, related-entity files, and the 3 most recent overlapping action items. The drafted body, thread context, and personalization signals are stored in the action file's ## Compose payload body section.
2. At click time, compose_view reads the action file, lifts the ## Compose payload YAML, and renders the iframe with the editable draft already shown.
3. You edit, then click Send. The compose iframe emits a committed envelope back to chat that targets the user's Slack Connector directly (Use the Slack Connector to send a Slack message as a thread reply. channel_id: …, thread_ts: …. Body: «…».). The host reads the envelope and dispatches to slack_send_message / slack_schedule_message / slack_send_message_draft with the channel_id + thread_ts inline in the envelope. No agntux-slack skill round-trip, no disk read.

No Slack write tool is ever called without a committed envelope from the iframe. The iframe Send button is the explicit authorisation gate. Action files written by 2.x.x sync runs (without ## Compose payload) surface the graceful compose_payload_missing error inside the iframe.

The sync skill runs inline in whatever context the host hands it (interactive chat or the scheduled-task scaffold) — no context: fork, no nested general-purpose agent, no frontmatter tools: whitelist. The skill inherits the parent's full tool surface (including the UUID-prefixed Cowork connector tools mcp__<uuid>__slack_*) and, critically, the parent's working-directory grant. The previous "router skill + sub-agent" and context: fork + agent: general-purpose shapes are both retired: each added a context boundary that did NOT inherit the host's "Allow for all scheduled runs" grant, so every scheduled fire silently re-prompted and the skill exited clean without advancing the cursor. See the Claude Code skill docs for the inline-skill shape.

UI handlers

Slack reply composer (ui://slack-compose): When you click Draft a reply or Schedule a reply on an action item, the host routes the click directly to the compose view tool, which lifts the pre-composed draft + thread context from the action file's ## Compose payload body section and renders the iframe inline. You see the channel name, the parent message, the last reply quote, and an editable textarea prefilled with the draft. Mode tabs let you choose Send now / Schedule / Save Slack draft. A "Why this draft?" disclosure surfaces personalization signals from user.md and per-plugin instructions. The Send button emits a self-contained envelope addressed at the user's Slack Connector (Use the Slack Connector to send a Slack message as a thread reply. channel_id: …, thread_ts: …. Body: «…».). The host dispatches directly to slack_send_message / slack_schedule_message / slack_send_message_draft. The Send button is your explicit confirmation; no Slack write happens before it. Discard is a pure local action — it shows a "Discarded" banner and emits nothing to chat.

Slack canvas summariser (ui://slack-canvas): When you click Summarise to canvas on an action item, the host routes the click directly to the canvas view tool, which lifts the pre-composed canvas sections from the action file's ## Canvas payload body section. Four editable section blocks (TL;DR, Decisions, Open questions, Participants) plus an editable title come pre-populated. A Preview tab shows the assembled markdown. The Create button emits a Slack-Connector-targeted envelope (Use the Slack Connector in two steps: 1. Create a Slack canvas titled «…» … 2. Post it as a thread reply in channel_id: …, thread_ts: …). The host runs slack_create_canvas, takes the returned canvas URL, and posts it as a Slack mrkdwn link in the thread. Discard is a pure local action — it shows a "Discarded" banner and emits nothing to chat.

Both UIs ship as embedded component bundles (no external S3 fetch). Build and bundle both after any UI component changes via the top-level builder:

node scripts/build-plugin.mjs agntux-slack

Or, when iterating locally with MCPJam Inspector:

node scripts/build-plugin.mjs agntux-slack --serve

Limitations

• Reads only. The sync skill writes nothing back to Slack. The draft skill is the only path that calls Slack write tools, and only after explicit user confirmation.
• DMs and group DMs are covered. Multi-party DMs (mpim) work the same way as channels via the per-channel cursor map.
• Bot messages are skipped by default. Add bot_id:<id> rules to data/instructions/agntux-slack.md → # Always raise to allow specific apps.
• Volume caps: 200 messages per channel per run, 10 action items per run. Hot threads update existing action items rather than spawning duplicates.
• Tracked-threads registry evicts threads with no activity for 30 days. New replies on long-dormant threads are caught via the discovery search (if you're @mentioned) or by re-discovery via the channel cursor (if the parent is touched).

Hooks and MCP server

AgntUX Slack ships no hooks/ directory. The Slack data connector is host-installed (declared via requires_source_mcp: { source: connector, connector_slug: slack } in marketplace/listing.yaml); the MCP server in mcp-server/ is the plugin's own MCP App UI server (compose / canvas view tools), not the data connector.

License

Apache License 2.0. See the LICENSE and NOTICE files at the repo root for full terms.

Support

• Bugs and proposals: https://github.com/AgntUX/AUX-plugins/issues?q=label%3Aagntux-slack
• Email: support@agntux.ai