Use the Slack Agent

​

Use the agent in Slack

• Direct messages: open the agent from your apps in Slack and write in its Messages tab. It answers every message.
• Channels: invite the agent to a public or private channel, then mention it, for example @Phaseo which model should we use to summarise call transcripts?. It replies in a thread. Mention it again for each follow-up in that thread.

• It reads the newest part of the thread, up to your message, so it can follow the conversation. Slack lets it read 15 messages at a time, and lets Phaseo’s app read about once a minute in each Slack workspace.
  • It always sees the thread’s first message.
  • In a direct message, it reads the latest messages of your conversation with it.
  • In a channel, it reads the latest messages from the last 30 minutes at most.
  • When more messages were written than it could read, it’s told that some are missing.
  • When Slack doesn’t let it read the thread just then, it works from its own earlier replies in the thread.
• It answers one message at a time in each thread. Up to 5 more messages wait their turn; after that it asks you to wait for the earlier answers.
• Slack’s stop button stops the answer it’s working on.
• Each answer ends with the model and what the answer cost in model tokens, such as Auto → GPT-6 Astra · $0.0031. With Auto, the model it picked follows an arrow, and other models the answer used follow a +. A model’s name links to its page on Phaseo, and the cost links to that answer’s requests in Logs, which members of the Phaseo workspace can open.

​

Choosing a model per message

@Phaseo with Claude Opus 5.5: check this incident summary for gaps
with openai/gpt-6-astra: draft a reply to this customer

• Only the allowed models are matched, ignoring case. The start of a name or ID is enough when only one allowed model has it, so with claude: works when one Claude model is on the list. An ID also works without its lab, such as with gpt-6-astra:.
• The agent answers that message with that model, without the with …: part, and the footer names the model that answered.
• If you name a catalogue model that isn’t on the list, or the name fits more than one, the agent replies privately with the models you can choose and doesn’t answer. That message isn’t counted as an answer. A message that starts with “With …:” but names no model, such as “With regards to the invoice: …”, is answered as usual.
• Each answer keeps the same limits, $0.50 and 12 steps, whichever model answers. The model’s tokens are billed at its own price.
• The Slack agent guardrail still applies. If no provider of the chosen model meets it, the agent says so instead of answering.

​

Asking another model

@Phaseo draft a reply to this customer, then ask Claude Opus 5.5 to check it

• The other model only sees what the agent sends it. It doesn’t see the Slack thread, and it can’t search the web or use tools.
• The agent can ask other models up to 3 times in one answer. Their cost counts toward the same $0.50 limit, and the agent stops asking when too little is left to finish its answer.
• The footer names every model used, such as GPT-6 Astra + Claude Opus 5.5 · $0.0123, and the cost includes them.
• If a model can’t be used, for example because none of its providers meets the Slack agent guardrail, the agent is told why and answers without it.

​

Comparing models

@Phaseo compare GPT-6 Astra, Claude Opus 5.5: what are the risks of moving our database to another region?
compare gpt-6-astra, claude, auto: summarise the EU AI Act in five bullet points

• Models are matched like with <model>:, only from the allowed list. If a model isn’t on the list, a name fits more than one model, you name fewer than 2 or more than 3, or there’s no question after the colon, the agent replies privately and doesn’t answer. That message isn’t counted as an answer. A message that names no model at all, such as compare the two plans, then pick one: …, is ordinary text and the agent answers it as usual. While model choice is off, a compare …: message that names catalogue models gets a private reply that Compare is off, and isn’t counted as an answer.
• Each model gets your question, not the rest of the thread, and can search the web and read web pages. They answer at the same time.
• The reply has a section for each model: its name, linked to its page on Phaseo, what it cost, and its answer. Each model is asked to keep its answer short, and long answers are cut so they all fit in one Slack message.
• If a model fails, its section says why, and the other answers are still shown.
• A comparison counts as one answer. The [$0.50 limit](/v1/guides/slack-agent-admin#limits-on-each-answer) is split evenly between the models, and the footer names them all and shows the total. The split only decides whether a model is asked: one isn't asked when its share is under $0.02. An answer isn’t stopped at its share, so it can cost more; each model’s reply is capped in length instead.

​

Questions with buttons

• Only the person it’s answering can answer. Anyone else who selects a button gets a private note.
• You can also reply in the thread instead (in a channel, mention the agent). Your reply is taken as your answer.
• Once you answer, the buttons are replaced with your answer and the agent carries on from where it stopped.
• A question waits for up to an hour. After that, the question says it expired and the answer ends there. Ask again when you’re ready.
• Slack’s stop button also works while it waits.
• While it waits, other people’s messages in the thread wait their turn, and they’re told privately who it’s waiting for.
• One answer can ask up to 3 questions. The whole thing counts as one answer, and the limits on each answer cover all of it, before and after each question.

​

Check the Home tab

​

Charts, images and files

• Charts: ask for a chart, for example @Phaseo chart our spend by week this month. The agent draws a bar, line or pie chart from the numbers it has and posts it in the thread as an image. Chart text uses Latin and Cyrillic letters; anything else stays in its reply.
• Images: ask it to make one, for example @Phaseo make an icon of a paper plane in blue. It uses GPT Image 2 through the agent’s API key, billed like any other request. An image costs a few cents and counts towards the answer’s [$0.50 limit](/v1/guides/slack-agent-admin#limits-on-each-answer), and the answer's footer includes it. One answer makes at most 2 images, and only while at least $0.15 of its limit is left.
• Files you share: attach images (PNG, JPEG, GIF or WebP, up to 5 MB each) or PDFs (up to 10 MB) to your message, or to earlier messages in the thread that the agent reads. It reads up to 4 files, 12 MB together, for each answer, newest first. Other file types can’t be read yet.

​

Searching Slack

What did we decide about the launch date?
@Phaseo find the latest incident review in #ops and summarise it
Who has been working on the billing migration this month?

• It searches as you. It uses Slack’s real-time search with the permission Slack gives it for your message, so it finds only what you can see in Slack. It searches public channels only, never private channels, direct messages or group DMs.
• It links what it uses. Each message or file it relies on is linked, so you can open it in Slack. When a search finds nothing, it says so.
• It can narrow a search to a channel, a person or a date range, and searches up to 4 times for one answer, 10 results at a time.
• By meaning or by keyword. Questions in your own words are searched by meaning on Slack plans that include AI search (Business+ and Enterprise+); other plans search by keyword.
• Only providers that keep no data. Once an answer has searched Slack, the rest of it, including another model it asks or an image it makes, goes only to AI providers with zero data retention, so what it found never sits in a provider’s logs. Web search and reading web pages are also off for the rest of that answer, so nothing it found can leave in a search or a link: ask it to search the web first if you need both.

• It searches only when answering someone’s message: Slack gives it permission to search with each message that mentions it or is sent to it directly, and the agent keeps that permission only until it has answered.
• Slack doesn’t let apps search for guests, and the agent never searches for people from other companies.
• Slack lets each person search about 10 times a minute. When it’s busier than that, the agent answers with what it has and says so.
• Slack allows this search only for apps in the Slack Marketplace and for a company’s own apps. Phaseo’s Slack app can search once its Slack Marketplace listing is approved; until then, the agent says search isn’t available. To search now, use your own Slack app.
• Installations from before search need updated Slack permissions.

​

Ask about usage

• requests and spend over a period, by day, week or month, and by model or provider;
• the credit balance, credits held for requests in progress, and credits spent this month;
• requests and spend for each API key;
• models in the Phaseo catalogue.

​

Memory and skills

• @Phaseo remember that deploys freeze on Fridays after 3pm
• @Phaseo pin this for everyone: our fiscal year starts on 1 April
• @Phaseo add a to-do: send the Q4 budget to finance, and later @Phaseo the budget to-do is done
• @Phaseo what do you remember about deploys? or @Phaseo forget the note about deploys

​

Skills

​

What admins control

• Let the agent keep notes (on by default, under Memory and skills in Settings → Slack) turns notes, pins and to-dos off. While it’s off, the agent saves none and sees none, and tells people who ask it to remember something that an admin has turned memory off. Skills keep working. The notes already saved are kept, and owners and admins can still see and delete them; turn it back on and the agent uses them again.
• Each Slack workspace keeps up to 500 notes and 100 skills. When memory is full, the agent says so; delete some to make room.
• Owners and admins see everything the agent has saved under Memory and skills in Settings → Slack. They can filter notes by kind, delete any note or skill, and edit a skill’s name, description and instructions. The workspace activity log records these changes without the text.
• Uninstalling the agent deletes all its notes and skills. Adding it again starts with an empty memory.

​

Scheduled tasks

• @Phaseo every weekday at 9:00, summarise yesterday's news about electric vehicles
• @Phaseo every Monday at 8:30, list last week's Phaseo spend by model
• @Phaseo on the 1st of every month at 10:00, list the API keys nobody used last month
• @Phaseo which recurring tasks are set up here? or @Phaseo cancel the Monday spend summary

• How often: every hour (at the minute you choose), every day, on weekdays (Monday to Friday), once a week on a day you choose, or once a month on a day you choose (in shorter months, on their last day). Never more often than once an hour.
• When: at a local time in the Slack time zone of the person who set it up, also when the clocks change.
• Where: in the channel or direct message where it was set up. Each run posts a short Scheduled: … message there and answers it in the thread. Replies in that thread are ordinary messages to the agent.
• What runs: the agent saves the request as a self-contained instruction, so each run works without the conversation it was set up in. Scheduled runs don’t ask questions with buttons or search Slack, and can’t set up or change recurring tasks.

• If the agent has been paused in Phaseo, or a Phaseo bill is unpaid (Payment due), the run is skipped and the task carries on: it runs again at its next time once the agent is back on, with nothing to resume. These skipped runs don’t count as failures, and Settings → Slack shows why the last run was skipped.
• If the person who set it up can no longer use the agent, or the agent can’t post in the conversation any more (for example, it was removed from the channel), the run is skipped and the task is paused, with the reason in Settings → Slack.
• Any other run that doesn’t go through, such as one refused because the Monthly spending cap is reached, counts as a failure. After 3 in a row the task is paused, and the agent tells the person who set it up in a direct message.

​

Recurring task settings

​

Shared pages

• @Phaseo make a page comparing our spend by model this month
• @Phaseo turn the decisions in this thread into a one-page summary

​

Page limits

• A page is kept for 30 days by default, then deleted. Admins can choose 7, 30 or 90 days for new pages.
• Each answer can publish up to 3 pages, and each Slack workspace keeps up to 200 pages at a time. A page can be up to 512 KB.
• Publishing is part of the answer: it costs nothing extra, and the model tokens spent writing the page count toward the answer’s $0.50 limit.
• People from other companies in shared Slack Connect channels can’t publish pages.
• An answer that has searched Slack can’t publish a page, because a page keeps what the agent wrote (see Searching Slack).

​

What admins control for pages

• Let the agent publish pages (on by default) turns publishing off. Pages already published stay until they expire or you delete them.
• Keep new pages for 7, 30 or 90 days.
• Anyone with the link can view pages (on by default). Turned off, every page needs a Slack sign-in, including pages already published.
• New pages are open to sets the default when both options are available.
• The list shows each page’s title, who asked for it, when it was published and when it expires, and how often it was opened. Owners and admins can delete any page, and its link stops working at once. They can open a page when its access lets them: a page for the Slack workspace needs them to sign in with Slack as a member. The workspace activity log records deletions without the title.