Short answer. Create a Slack app, enable the Events API, subscribe to
app_mention, addapp_mentions:readandchat:write, and run a Bolt app. Reply in the thread withthread_ts, deduplicate byevent_id, and call the model asynchronously after Slack has been acknowledged.
1. Create the Slack app
- Create an app at
api.slack.com/apps. - Enable Socket Mode for local development, or the Events API with a public HTTPS request URL for production.
- Subscribe to
app_mention(andmessage.channelsonly if you need ambient messages). - Add bot scopes:
app_mentions:read,chat:write. Addchannels:historyorgroups:historyonly if you read thread context. - Install the app and store the bot token and signing secret in a secret manager.
2. Minimal Bolt app
import os
from slack_bolt import App
from slack_bolt.adapter.socket_mode import SocketModeHandler
from openai import OpenAI
app = App(
token=os.environ["SLACK_BOT_TOKEN"],
signing_secret=os.environ["SLACK_SIGNING_SECRET"],
)
client = OpenAI(
api_key=os.environ["PLUGSKY_API_KEY"],
base_url="https://api.plugsky.com/v1",
)
@app.event("app_mention")
def handle_mention(event, say, client):
thread_ts = event.get("thread_ts") or event["ts"]
question = event["text"]
answer = ask_model(question)
say(text=answer, thread_ts=thread_ts)
def ask_model(question: str) -> str:
response = client.chat.completions.create(
model="plugsky-plus",
messages=[
{"role": "system", "content": "Answer concisely. Say when you are unsure."},
{"role": "user", "content": question},
],
temperature=0.2,
)
return response.choices[0].message.content
if __name__ == "__main__":
SocketModeHandler(app, os.environ["SLACK_APP_TOKEN"]).start()
For HTTP mode, use the Bolt request handler for your web framework and mount it at the Events API URL. Socket Mode is for development; production usually uses HTTP.
3. Acknowledge within three seconds
Slack expects an acknowledgement quickly. Bolt acknowledges the event before your handler runs, but a slow handler still blocks the worker. For a model call, enqueue the work and reply with chat.postMessage when it finishes. If the model takes 20 seconds, the user should see a “working on it” message, not silence.
4. Thread context
Use conversations.replies with the thread's ts to fetch prior messages, then trim to the last N turns and a token budget. Do not resend the entire channel history. Strip bot messages and system noise, and keep user IDs for attribution.
5. Retrieval
context = search_knowledge_base(question, top_k=5)
prompt = [
{"role": "system", "content": "Use only the provided context. Cite the source URL."},
{"role": "user", "content": f"Context:\n{context}\n\nQuestion: {question}"},
]
If no permitted source is found, say so and offer to escalate. Do not let the model invent a policy answer in a workspace where people will act on it.
6. Deduplicate events
Slack retries event deliveries. Store event_id (or client_msg_id) with a short TTL and ignore duplicates before calling the model. Also check whether the bot already replied in the thread; a duplicate reply is worse than a missed one.
7. Escalation
- Low confidence or no retrieval result.
- Requests involving billing, refunds, permissions or security.
- Repeated follow-ups from the same user.
- The model provider is rate-limited or unavailable.
Escalate with a thread reply that names the human queue and includes the question, sources and confidence. Do not silently stop replying.
8. Rate limits and retries
Slack returns 429 with Retry-After; the model provider does too. Queue the job, respect the delay, and cap retries. If the model is down, post a short fallback message and create a ticket instead of retrying forever.
FAQ
How do I build a Slack AI bot?
Create a Slack app, enable the Events API or Socket Mode, subscribe to app_mention, and run a Bolt app that calls your model API.
Which scopes do I need?
app_mentions:read and chat:write at minimum; add history scopes only if you read thread context.
Why does Slack say the app did not respond?
Slack expects an acknowledgement within 3 seconds. Acknowledge first, do the model call asynchronously, then reply.
How do I stop duplicate replies?
Store event_id and ignore retries; check whether the bot already replied in the thread.
How do I handle rate limits?
Respect Retry-After for Slack and the model provider, queue work, cap retries and post a fallback.
Reviewed 20 November 2026. Slack scopes and Bolt APIs vary by version; check the current Slack documentation. Keep a human escalation path for permissions, billing and security requests.