GPT DotFIELD GUIDE
Home/Integrations
Integrations

How to add your ChatGPT Dot to Slack: setup, admin approval, reply times, and what to do when replies never arrive (OpenAI Dots)

Open your Dot's profile, choose Add and connect Slack, then DM it or mention it in a thread; on Enterprise an admin must allow dots in Slack first. Mention it again in every thread reply, test in a DM first, and don't remove the connection to rename it. Real ChatGPT Dots (OpenAI Dots) Slack setups, MCP -32603 reply failures and the fixes that worked.

GPT Dot · Updated

The short version: what works in Slack, how to set it up, and where replies get stuck

Should you connect your Dot to Slack?

Yes, if you want to hand it jobs from your phone, get scheduled reports in a channel, have long threads read for you, or keep a document current with what your team decides in Slack. Reading and reporting work well: Seowoo Han gets a daily briefing in the channel set aside for it (case 1), happytlog receives scheduled results while the computer is off (case 2), Pat Simmons keeps one thread per job on the way home (case 6), and Zinho Automates' Dot found where a long redesign thread was stuck (case 13). Nida Mohsin used a week of Slack conversations to keep a brand FAQ, a project deck and a research plan up to date, and her Dot only brought her the updates that needed her attention, approval or decision. The talking-back side is weaker. Toby's DM to a colleague failed and a group request waited about 50 minutes (case 18), channel replies have failed with MCP error -32603 while DMs worked (case 20), a follow-up in a channel thread is ignored unless you mention the Dot again, several people couldn't add the Dot back after removing it (case 21), and calling the Dot from Slack failed for Allie K. Miller (case 28). It suits people who already work in Slack and can wait a few minutes for an answer.

How to set it up

  1. Check your plan and workspace. A Dot comes with Pro ($100, $200 or $500 tiers) outside the EEA, UK and Switzerland, with Business Premium in any supported region, and with Enterprise once an admin switches it on (How to get ChatGPT Dots; in Europe, see Dots in the EU and UK). On Enterprise, a workspace owner has to turn on Use dots (Beta) and Add dots to Slack and Microsoft Teams, and if your Slack workspace requires app approval, a Slack owner or app manager has to approve the install (Manage dots in ChatGPT workspaces). The same permission covers Microsoft Teams.
  2. Settle the Dot's name and the Slack account before you connect. Fotios removed the Dot to rename it and Marcroberts removed it to move it to another account, and neither could add it back (case 21).
  3. Open your Dot's profile, select Add and connect Slack (Message your dot). Read the permission screen before you allow it: uehara's asked for scopes such as posting as him, reading DMs and creating channels, and the Dot then introduced itself by DM (case 27). Apps such as Gmail or Drive, and your computer, are connected separately.
  4. Test in a DM before you rely on a channel. Send one small read-only job and open the thread, because uehara's answer came after 81 seconds as a thread reply (case 27). Then add the Dot to a channel, mention it in a thread, and ask it to read the channel back after it posts (case 20). When WEEL tried this right after connecting, the first channel mention got no reply; sending it again a little later worked.
  5. In a channel thread, mention the Dot in every reply. loweaxerium found that after the Dot answers or asks a question in a thread, your next reply reaches it only if it mentions the Dot again. When a job needs a real back-and-forth, move it to your DM with the Dot.
  6. Say who it works for and what it may share. By default it responds only to you, and other people's DMs or mentions don't start work for your Dot. You can tell it to engage with others; before it shares something from a private conversation, it checks that you've allowed it (Message your dot). Everything it posts in a channel is visible to the whole channel.
  7. Say where each kind of update goes. Seowoo Han gave every job its own channel (case 1), 단테랩스 named chat plus Slack together with the time (case 4), and OpenAI suggests keeping routine progress in ChatGPT and bringing decisions to you in Slack.
  8. Make every watch explicit. Adding the Dot to a channel doesn't start a watch. Name the channels, what to look for and when to tell you, then ask it to confirm the recurring task, as まさやん did with a check every 10 minutes (case 3). A forwarded message with "Keep an eye on this" sets the topic, and the channels and trigger still need saying (case 5).

Where it gets stuck, and what worked

  • It reads the channel but can't reply, with MCP error -32603. nittajun's channel replies came back on 4 October once the Dot used its dedicated Slack sending route, while the general route still failed for markguoz on 5 October and for another user on 9 October (case 20). What worked: ask which route it used, have it read the channel back, and use a DM to yourself as the quick test.
  • Your follow-up in a channel thread gets no answer. Without a fresh mention, your reply doesn't reach the Dot, so a question it asked can sit unanswered on both sides (loweaxerium, 18 October, who has asked OpenAI for an opt-in to keep one thread open). What works: start each reply in the thread with the Dot's mention, or carry the conversation on in a DM.
  • The first mention after connecting gets nothing. WEEL's first channel mention went unanswered right after setup, and the same message sent again a little later got a reply. Wait a few minutes after connecting and send once more before you suspect the connection.
  • It says it's done before the message has gone out. Toby announced the Slack connection before checking it could send the DM Mark Nguyen wanted, and the group message only went out on an explicit route (case 18). uehara's mention summary sat unsent after "I'll check" until he nudged (case 27). What worked: ask for proof that the message arrived, and after a few minutes of silence ask whether the result was sent.
  • The answer is in a thread you didn't open. uehara's DM list showed only "1 reply" (case 27), and ishizakahiroshi's Dot answered inside each request thread, so the top-level messages looked silent (case 24). Open the thread before you ask again.
  • It's slow, or needs pushing. uehara's DM got an acknowledgement in 13 seconds and the answer in 81 (case 27), but a colleague's PDF request was picked up after about 50 minutes (case 18), and Mark Kashef kept having to push the Dot to act (case 19). Use Slack for work that can wait, and ask the Dot to say when you can expect the answer.
  • It won't come back after you removed it. Fotios, Marcroberts and scottinallcaps all got stuck at "Slack setup failed" or a generic error, and for scottinallcaps the Dot-specific Slack app was never created (case 21). What worked for Fotios: a handover document written into a Space, then deleting the Dot and creating a new one from ChatGPT on the web. That route has its own risk, because ewerton.lourenco saw "Failed to load your dot" for a day after deleting. Marcroberts kept Slack work going without recreating anything: the Dot uses the Slack plugin, interactively and on a schedule, so it posts under the plugin account's name and only sees mentions at its next scheduled check. The safer fix is not to remove the connection in the first place.
  • A post may go out twice. Chris Eberl found that an accepted API request isn't proof that a post is live, so he checks the publishing status before any retry and approves one exact version per thread (case 22).
  • Another bot ignores the Dot. 銀河's other bot dropped the Dot's first test as a post from a bot; adding a bot allow-list on that side fixed it, and the retest got a reply about 19 seconds later (case 26).
  • A custom client is minutes slow. Through MCP Events, shiyubinguzhen measured about 270–303 seconds from send to visible reply on simulators, and hjotha got replies in about 10 seconds by forwarding through Slack instead (case 29). Time your own route before you build on it.

For calls from the ChatGPT app, see the voice calls page; for other errors, start at Dots not showing or not working.

What it can and can't do yet

What it already does well:

  • Posting a daily report to the channel set aside for it (Seowoo Han), or a morning summary to chat and Slack at once (단테랩스).
  • Delivering scheduled results while your computer is off (happytlog).
  • Taking jobs in Slack while you are away from your desk, with one thread per job (Pat Simmons, Mustafa Ergisi).
  • Reading a long thread and saying who is waiting on whom (Zinho Automates).
  • Keeping an FAQ, a deck or a plan in step with what the team agrees in Slack, and passing on only what needs you (Nida Mohsin).
  • Noticing things in channels where nobody tagged it (Rajat, GENEL's quoted user).
  • Adding the Slack history with attendees to a meeting brief (AI 서대표).
  • Producing a file in a format ChatGPT doesn't export, once the tool is installed on its computer (tangokorea).
  • Waiting for your yes in a thread before it publishes (Chris Eberl, 坂本匡史).
  • Talking to another bot in the same thread once that bot allows it (銀河).

What it doesn't do yet, or not reliably:

  • Answering in real time; expect a minute or two, sometimes much longer (cases 18 and 27).
  • Posting through every sending route; some channel replies fail with -32603 (case 20).
  • Picking up your next reply in a channel thread unless you mention it again.
  • Taking work straight from teammates' DMs or mentions, since only you direct it.
  • Being added back after the Slack connection was removed, for some users (case 21).
  • Answering a call started from Slack (case 28).
  • Watching a channel just because it was added; a watch needs a schedule (cases 3 and 5).
  • Showing in Slack what you said to it in ChatGPT; each Slack conversation shows only the messages exchanged there, though the Dot can use what it learned elsewhere.

One prompt to copy

Built from what worked for nittajun, uehara and Seowoo Han, and written to catch the delivery problems above on day one:

You're now connected to Slack. Before I rely on it, run a quick check. Reply to this DM in its thread, then post one line in #[test channel] and read the channel back to confirm it's there. Tell me which sending route you used, and if anything fails, quote the exact error instead of retrying silently. Then confirm the rules: you take work only from me, you ask me privately before sharing anything I've told you in private, and you don't post outside #[channels] without my written yes. If a channel thread needs more than one answer from me, ask me in this DM instead, because my replies in the thread only reach you when I mention you. After that, set up this job: every weekday at [time] [timezone], post [what] to #[channel], and confirm the schedule with me before the first run. If you go quiet for more than [N] minutes after saying you'll check, tell me where it stopped.

Prompts for single jobs and problems are under More copyable asks at the bottom.

Reports and watches posted to Slack

1. One channel per job, with a daily report posted there (Seowoo Han)

Seowoo Han (@swhan0329_) connected Slack and told the Dot, in plain Korean, to "research the market every day and post a report to this channel." On 1 October the #news channel got a structured "AI · Computer Vision · Physical AI briefing" timed 09:56 KST, with five items and links. Because one Dot handles everything, Han created a separate channel for each purpose (market research, reference research and so on) and told the Dot where each daily report belongs, which keeps conversations sorted by topic even with several jobs running.

Every weekday at [time] [timezone], research [topic] and post a short report to #[channel]: five items, one line each on why it matters, with a source link per item.

2. Scheduled results delivered to Slack, even with the computer off (happytlog)

Korean developer happytlog set up several scheduled tasks, then asked the Dot whether it could chat in Slack and send the scheduled results there as notifications. The Dot replied with a Slack connection link and the setup was done right away. Since then happytlog assigns work through Slack while the computer is off.

3. A channel checked every 10 minutes, with a "don't send email" rule (まさやん)

Japanese creator まさやん gave a Dot two recurring jobs: deliver the day's calendar as an image every morning at 7, and watch a Slack channel every 10 minutes. According to the video description, the Dot suggests actions from connected app data without being asked but asks before sending or deleting. まさやん added one custom rule ("don't send email") and chose which apps it may use in the plugin settings.

4. A morning email summary sent to both chat and Slack (단테랩스)

Korean channel 단테랩스 set up a scheduled task during a 30-minute phone call with their Dot: every morning at 7, summarize only the important emails from the last three days and send the summary to the ChatGPT chat and to Slack. The same video keeps in the parts that stalled: a Coupang product search was blocked on both the local and the cloud computer, and the Dot once misreported its progress before correcting itself.

5. Forward a message and say "Keep an eye on this" (Alexander Embiricos, OpenAI)

Alexander Embiricos, who works at OpenAI, shared a three-step prompt on 2 Oct (mirrored by vibecoding.tech): spot something important in Slack, click Forward to send it to the Dot, and write "Keep an eye on this." The forwarded message sets the topic. If you want a regular check rather than a one-off look, add which channels to watch and what change should trigger a message.

Giving jobs from Slack

6. One thread per job, on the way home (Pat Simmons)

The mobile app didn't show Pat Simmons's Dot the night Pat left the office, so Pat connected László to Slack and kept giving it jobs on the way home. Each job got its own thread and the replies stayed organised; Pat still prefers this to the mobile app. While shopping it sent a screenshot and asked "green or blue?" in Slack. László also found Code Queue, the Linear-style board Pat's agents use, and scheduled a status summary in Slack every weekday at 9 PM. On a later call it reported 137 issues in progress and 122 in review, and Pat asked it to propose which stale ones to close, send the list in Slack and close them after approval. About a minute after the call ended, the "Code Queue cleanup" message arrived.

7. A solo Slack workspace as a personal work archive (이완성)

이완성 (@autowanseong) had work scattered across many Codex sessions, so 이완성 started gathering it in a Slack workspace used by one person: internal project progress sheets, email summaries and video material. The Dot is asked to find and organize them, and the next task picks up from the Dot's post.

8. A game app built during a flight (Mustafa Ergisi)

On a flight, Mustafa Ergisi (@mergisi, 2 Oct) asked the Dot in Slack to build something fun for the plane. About 20 minutes later the game app was playable, and the session was recorded.

9. A design document as an .hwp file (tangokorea)

Korean user tangokorea (5 Oct) worked through Slack: the Dot installed rhwp, an open-source tool for Hangul word-processor files, on its cloud computer, and was then asked to produce a design document as an .hwp file. It did. If your team needs a file format ChatGPT doesn't export, this is the pattern: get the tool onto the Dot's computer first, then ask for the file.

10. Coordinating an AI marketing team from Slack (Grace Leung)

In a 3 Oct build video, Grace Leung sets up brand context, reusable marketing skills and specialist agents in ChatGPT, then messages the Dot on Slack so it coordinates that team on a campaign while Grace is away from the desk. The video carries a HubSpot sponsor link; the build itself is Grace's own.

11. A 45-minute call that cleared the day's email and Slack tasks (Riley Brown)

Riley Brown (1 Oct) called the Dot for 45 minutes, phone in pocket and AirPods in, and "handled all my email / slack tasks for the day." In a later video, Riley calls the Dot, Bluey, hands-free; it sends emails and Slack messages and then starts a local Codex session to redesign a landing page.

Reading threads and noticing things

12. A Slack scan in the very first session (happytlog)

In happytlog's first session, the Dot went through their Slack messages while its name was still being chosen. It found a task planned with a collaborator and offered to collect decisions scattered across Slack into a checklist for handing off to development. If that would surprise you, name the channels the Dot may read in your first message.

13. Where a long thread is stuck (Zinho Automates)

In a setup video, Zinho Automates had a Dot named Zozo read an entire Slack thread about a website redesign and report where the project was stuck and who was waiting on whom.

Read [thread link] and tell me where [project] is stuck: the open question, who is waiting on whom, and the last message from each person, with permalinks. Don't reply in the thread.

14. Noticing what nobody tagged it on (Rajat)

Rajat's Dot, Lex, had no prompts or routines set. Within 24 hours it noticed two cohort members drifting from what had been covered in a session, in a Slack channel where it wasn't tagged, and warned that Rajat's own changes would be delayed because a teammate was discussing a different set of changes on Slack.

15. Replies and an invoice handled in parallel (a user quoted by GENEL)

The Japanese account GENEL (@ai.genel) quoted another user's report: the Dot joined Slack conversations without being mentioned, flagged that the month's invoice hadn't been handled, then, when asked, replied to Slack messages, saved screenshots for an X post and finished the invoice, confirming each request as it was done. GENEL called a Slack agent with full context "far and away the best" of its kind.

16. Slack history before video calls, and a thread set up from a holiday (AI 서대표)

Korean channel AI 서대표 connected only Gmail, Google Calendar, Drive and Slack for a three-day test. Before video meetings, the Dot attached the Slack history with the attendees to its briefing. On holiday abroad and unable to make the next day's meeting, 서대표 only asked it to pass on what the team needed. The Dot started a Slack thread with the teammates, confirmed the meeting time and the decision that came out of it, reported back and saved the outcome to memory.

17. Teammates messaging the Dot directly (Justin Schroeder)

Justin Schroeder (29 Sep) had early access and says Slack is where Dots "really shines": the team can message the Dot directly, and it can take actions such as spinning up new panes. Under OpenAI's current rules, only the owner directs a Dot, so teammates' messages don't start work on their own; you can tell it to engage with them.

Where it went wrong

18. A DM that failed and a 50-minute wait (Mark Nguyen, The Slide Factory)

Mark Nguyen's first-day review of the Dot, Toby, goes through the Slack setup step by step. Connecting Slack worked, but Toby announced the connection before checking it could send the DM Mark wanted, and the message to a colleague failed on the route it tried; sending to the group through an explicit route eventually worked. When a colleague asked for a PDF summary at about 4:24 p.m. Pacific, Toby picked it up around 5:13 and answered at 5:15. A daily SEO report to Mark's private Slack was scheduled for 8 a.m. the next day but hadn't run by the end of the session. The review's advice: start with read-and-flag work, and don't treat a Dot in Slack as a fast-replying teammate yet.

19. Easy setup, but it had to be pushed (Mark Kashef)

In a hands-on review, Mark Kashef found the Slack setup straightforward, but kept having to push the Dot to act and struggled to tell whether it was working.

20. Can read channels, can't reply: MCP error -32603 (Jeremy Zhang, nittajun and other GitHub reporters)

Jeremy Zhang reported on 5 Oct that a Dot could read Slack channels but failed when replying, with MCP error -32603 and DMs that only worked some of the time (root post). The same failure is tracked in openai/codex #50164, opened by nittajun on 2 Oct: the Dot received the mention in a private channel and read it, but the reply failed with "Mcp error: -32603: MCP request failed", while replies started from the Dot's Slack DM went through. Making the channel public or making the Dot a channel manager didn't change anything. nihal-docket saw even a plain top-level "Delivery test" fail in a channel with only the owner and the Dot.

On 4 Oct nittajun reported channel replies working again through the Dot's dedicated Slack sending route, checked by reading the messages back in two channels. The general message-sending route still returned -32603 that day. markguoz's error on 5 Oct names that general route (user_message.send_message), and another user confirmed the same symptoms on 9 Oct. So if replies fail, ask the Dot which route it used and to read the channel back, and use a DM to yourself as the quick test.

21. Removed from Slack to rename or switch accounts, and it won't come back (Fotios, Marcroberts, scottinallcaps and a delete-and-recreate thread)

The rename problem bradynapier reported on 7 October (Slack assistant case 12) picked up more reports in the same OpenAI Community thread. Fotios (8 October) followed the same steps: the Dot worked in Slack, Fotios removed it in order to rename it, and it could not be added again. Marcroberts (10 October) got there another way: to move the Dot to a different Slack account, Marcroberts removed the Dot's Slack connection, disconnected the Slack plugin, connected it to the other account and approved the permissions. Both accounts own their Slack workspaces, and adding the Dot now fails with a generic "failed, try again" error.

On 16 October, scottinallcaps hit the same wall after a fuller cleanup: removed the ChatGPT app from both workspaces, removed and deleted the Dot-specific Slack app, uninstalled the Slack plugin in ChatGPT/Codex (Plugins tab and the Dot's own tools), and ran Add to Slack several times. Each attempt ended with "Slack setup failed". The ChatGPT app/integration always reappeared in the workspace, but the managed child app, the Dot-specific app ChatGPT should create, never did. That points at the managed-app creation step on OpenAI's side, not a missing Slack permission.

The same day, Marcroberts described a stopgap that keeps some Slack work going. The Dot can still use the Slack plugin, both in conversation and on a schedule, but it then posts under the plugin account's name rather than as itself, and it doesn't wake when mentioned, so it picks up mentions, posts and replies at each scheduled check. Several Slack accounts can be connected, even in one workspace, and the Dot can be told which one to use. A few hours later, after an app update, the Slack button was gone from Marcroberts's Dot profile.

Fotios got out of it by deleting the Dot and creating a new one, only from ChatGPT on the web, after first asking the old Dot to write a handover document into a Space. That route has its own risk. In a separate thread, ewerton.lourenco deleted a Dot on 7 October to start over and then saw only "Failed to load your dot" with no way to create a new one; garymjr saw "Couldn't load your dot" after doing the same. Creating a Dot worked again for ewerton.lourenco the next day without any change on their side. One reader, whose calls to the Dot had been failing for a day, decided not to delete it after reading the thread.

Approving and handing over work in a Slack thread

22. One Instagram draft per thread, approved before it goes live (Chris Eberl, PFP Labs)

Chris Eberl (blog, 6 October 2026) set up a Dot to prepare one still image with a German caption at a time. Each draft arrives as a new top-level post in a private Slack channel, with the actual image and the full caption together, and he approves it in that post's thread, from his phone if needed. His rule is that a yes covers only that exact image and caption, plus uploading the public copy Instagram needs and publishing it. A revised version goes back to the thread for a fresh yes. Publishing runs through Windsor's API, the public JPEG link comes from Creatify, the dated archive stays private in Google Drive, and Meta Ads is connected read-only. The first route, publishing through the browser, failed with a generic cross-posting metadata error; the Windsor route worked. Two test posts went live, and their links came back in the same thread. A recurring draft for Monday and Thursday at 8 AM Los Angeles time was saved, with its first run due on 12 October, so it hadn't run yet when he wrote. His warning for anyone copying the setup: an accepted API request isn't proof that a post is live, so check the publishing status and Instagram before retrying, or a slow response can turn into a duplicate post.

23. A weekly analysis asked for in Slack, with an approval halfway (坂本匡史 / Masashi Sakamoto)

坂本匡史 (note, 7 October 2026, in Japanese), an AI solopreneur in Fukuoka, asked the Dot in Slack to analyse the traffic to Sakamoto's note posts and update the weekly analysis in "note studio", a note-management system Sakamoto built. An approval request arrived partway through. Once Sakamoto allowed it, the Dot carried on, entered the analysis and the per-article results, and reported that it had reloaded the page after saving. Some number fields stayed empty because Sakamoto hadn't given it the data. The Dot didn't guess at fields where it lacked the dates it needed; it said which fields were left and completed them once Sakamoto sent the missing information. The same post describes the Dot starting Slack conversations itself when new email continued earlier work. Those follow-ups are on the project follow-up page.

24. Jobs from ChatGPT and Claude Code, all in one numbered Slack thread (ishizakahiroshi)

ishizakahiroshi (Zenn, 4 October 2026, in Japanese) sends jobs to the Dot from two places: ChatGPT with its Slack plugin, and Claude Code with Slack's official MCP plugin. Both post to the same Slack DM. Each job gets a number from a ledger kept on ishizakahiroshi's own machine and a parent message titled "#number job name", and everything after that stays in its thread. The message itself is short: the number, the repository, a link to the instruction file pinned to a specific commit, a link to the job's PROGRESS.md, the scope, where to deliver, and what the Dot should confirm when it accepts. ishizakahiroshi treats "sent" and "accepted" as different states and waits for the Dot's acceptance reply: which commit it read, whether the tools it needs actually work, and which branch it will deliver to. The Dot's replies landed in the request thread, so looking only at top-level messages makes it seem as if nothing came back. One thing didn't work: the Slack connector set up in claude.ai never appeared in Claude Code, for reasons ishizakahiroshi couldn't find, so the Claude Code plugin is the route that worked. How these jobs ran on the coding side is on the Codex orchestration page.

Checking what you said, and talking to other bots

25. A Slack message it flagged as inaccurate, then corrected (Derrick Choi, OpenAI)

Derrick Choi (8 Oct 2026), who leads Codex in APAC at OpenAI, posted that his Dot called him out for saying something inaccurate on Slack and then corrected it for him. The post doesn't say what the mistake was.

26. A Dot and another bot replying to each other in one Slack thread (銀河 / @ginga_noguchi)

銀河 (@ginga_noguchi) (12 Oct 2026, in Japanese, also on Threads), a web marketer, now has his GPT Dot and his main Grok Bot talk to each other inside Slack, so the loop runs from him to the Dot to that bot and back. The main bot hands jobs to his other Grok Bots and on to Cursor Projects, Codex, Claude Code, programs on his own machine, and Notion Workers in the cloud. His screenshot shows how the first test went. In a thread, he asked the Dot to mention the other bot (shown in Slack as "Grok CoS") as a send-and-receive test. The Dot worked out that its 07:54 test had reached the other side but was dropped there as "a post from a bot"; a bot allow-list was then added at 07:56 and the bot restarted at 07:57. On the retest, the Dot confirmed a reply from the other bot's own account about 19 seconds after it sent its message, and noted that nothing had gone out under 銀河's name. It also said it had not yet checked whether the other bot's reply alone would start the Dot working. 銀河 says his work now moves forward with just his phone and Slack.

Reply times, and a call that didn't connect

27. An 81-second DM answer that landed in a thread, and a summary that needed a nudge (uehara, DEV Community)

uehara (DEV Community, 10 October 2026) connected a Pro Dot to eight work accounts, Slack among them, from 7 to 9 October and checked every result against the real apps. Right after Slack was added, the Dot sent a self-introduction by DM that included a warning: in group conversations, everyone in the conversation can see its replies. Slack's permission screen asked for scopes such as posting as him, reading DMs and creating channels. He then DMed: "Tell me how many events I have left today, with titles only. Do not change any events." By Slack's own timestamps, the acknowledgement came 13 seconds later and the answer 81 seconds later, as a thread reply, so the DM list only showed "1 reply" until he opened the thread. A separate request to summarise his Slack mentions in three lines went quiet for five minutes after "I'll check"; when he nudged it, the Dot said the check had finished but sending the result had failed.

28. A channel and a daily team summary, but calling from Slack failed (Allie K. Miller)

In her setup and live demo video (YouTube, 8 October 2026), Allie K. Miller connected her Dot to a Slack workspace through a permissions prompt; the Dot was given a channel, and she set a recurring daily summary of the AI use cases her team shares. Calling the Dot from Slack failed with "couldn't connect the call". Her framing for the product is ongoing responsibilities such as "watch my Slack and tell me if meetings change", rather than one-off prompts. (Slack details as summarised by OpenClawDatabase.)

MCP Events latency, and a Slack-forward workaround

29. MCP Events took 270–303 seconds; Slack-forward brought replies to about 10 seconds (shiyubinguzhen, hjotha)

On 14 October 2026, shiyubinguzhen reported building a custom iPhone/Apple Watch client that talks to an existing ChatGPT Dot through MCP Events (not a separate model API). In two measured simulator tests (paired iPhone/watchOS simulators, not physical devices), the reply eventually appeared in the Watch simulator reply view, but send-handler entry → displayed reply took about 302.6 s and 270.3 s. The largest observed interval in both samples sat between event creation and the payload becoming visible in the Dot’s main conversation (~212–226 s); they asked for guidance to separate event delivery, scheduling/batching, agent/tool work and the return path, and did not attribute the whole delay to OpenAI queueing. hjotha replied the same day: instead of talking to the Dot with MCP Events, forwarding the message through Slack brought replies in about 10 seconds rather than 60+ seconds. Treat the Slack-forward as one user’s workaround, not an official SLA.

More copyable asks

Every weekday at [time] [timezone], post to #[channel] a summary of [project]: decisions, open questions and owners, each with a permalink. Confirm the schedule before the first run.

I'm forwarding this message. Keep an eye on [topic] in #[channel-1] and #[channel-2], and DM me only if [condition]. Confirm what you'll check and how often.

Read the last [N] days of #[channel] and list requests addressed to me that nobody answered, with the person, the ask and a permalink. Don't reply to anyone.

When you post or reply in a Slack channel, read the channel back afterwards and tell me if the message isn't there. If a send fails with -32603, tell me which sending route you used, don't retry it silently, and DM me the text instead.

Before you change anything about the Slack connection:

Before I rename you, switch your Slack account or recreate you, write a handover document in the [Space name] Space: your name and Slack handle, the channels you're in, every scheduled task with its time and destination, the custom rules you follow, and the open jobs with their links. Tell me when it's saved.

Approving posts in a Slack thread:

For each draft, start a new post in #[channel] with the finished [image or file] and the full text together. Wait for my approval in that post's thread. A yes covers only that exact version; if anything changes, post the new version in the same thread and wait again. After publishing, check that it is actually live before you tell me, and put the live link in the same thread. If you're not sure it went out, check first and don't publish it again.

Testing whether your Dot and another bot can talk in Slack:

I'm testing whether you and [other bot] can talk in Slack. In this thread, mention @[other bot] once and ask it to reply once in the same thread from its own account. Tell me when you sent it, when the reply arrived, and which account posted it. If nothing comes back within [5] minutes, tell me where it stopped. Don't post anything under my name.

Related

Copy this task brief

Yes, a ChatGPT Dot works in Slack for reports, watches and jobs that can wait. How to connect it, what admins must allow, why thread replies need a fresh mention, where replies get stuck and what fixed it, plus a day-one check prompt.
Download Markdown ↓

Sources & further reading