Slack

Mention the GOAT bot in a Slack channel or send it a direct message, and GOAT answers in that Slack thread, starts the work in its own coding thread, and posts the result back to the same thread. The conversation stays in Slack; the GOAT Bot screen shows only a chip for the work and its final report.

Set up

Create a Slack app with Socket Mode turned on. GOAT opens the connection itself, so the app needs no public URL.
Give it an app-level token with connections:write, and a bot token with chat:write, app_mentions:read, and im:history. Subscribe the bot to the app_mention and message.im events. Turn on Interactivity — Socket Mode delivers button presses, so the app needs no Request URL — then install the app to your workspace.
In GOAT, open Settings → GOAT Bot → Slack for the machine that should do the work. Paste both tokens and turn on Use Slack. Tokens stay on that machine and are never shown again. The bot answers mentions in every channel you invite it to; choose These channels only under Channels to limit it to a list of channel IDs.
Invite the bot to each channel where people should mention it (/invite @your-bot). Slack does not send the bot mentions from channels it isn’t in, so a mention there gets no answer.
Optionally, under Link your Slack account, create a one-time code and send it to the bot in a direct message as link <code>. Linked people work at the level chosen for their code; everyone else in the workspace works at Access without a link (by default, Ask before changes).

Good to know

In a channel, mention the bot in each message; it does not read other channel messages. Direct messages always work.
Who can use it decides who may start work. Everyone in the workspace (the default) answers anyone in the connected Slack workspace; people from another organization in a shared channel are ignored. Linked accounts only answers everyone else with how to link. Someone without a link sees only their own work and the conversation they write in, and cannot use GOAT memory or change GOAT Bot’s notes.
Replies, questions, and results go to the thread where you asked, and questions tag the person who asked.
Default project decides where work from Slack runs when your message doesn’t name a project. Without it, GOAT picks the project from the conversation and asks only when it can’t tell, suggesting the one you worked in most recently. This machine (no project) runs work that isn’t about a repository.
GOAT Bot answers questions about the machine it runs on, such as its name, OS, memory, and what is running, without starting work.
For questions that need a look around but no code change (what fills the disk, what several threads concluded, something on the web), GOAT Bot says it is on it and looks into it itself, then answers in the same conversation. Messages you send while it works are added to what it is doing. It uses the model shown at the top of the GOAT Bot screen, which you can change there or under Settings → GOAT Bot (by default, the default model for new threads). Turn off the model switch in Settings to follow the default again. It runs with full access, never more than the person who asked may use, so it can read files, logs, and status anywhere on the machine without asking. It is told to change nothing itself and to start a coding task or ask you when something needs changing. That rule is part of its instructions; GOAT does not block changes, so link people at a lower access level if they should not be able to make the bot change things.
GOAT Bot and work started from Slack never see or control this computer’s screen on their own, even when computer use is on. When a task needs it, GOAT asks for approval in that task’s thread, for a set time. Someone allowed to change GOAT’s settings approves it in the GOAT app on desktop, web, or mobile; replying in Slack does not approve it. Approval admits that thread until the time shown, not each action, and you can end it early from the request.
Ask GOAT Bot about earlier work (“did the login PR merge?”, “what did the docs thread do?”, “when did I mention the lighthouse score?”) and it looks it up itself and answers, without starting a coding agent. It reads the work it started, the pull requests that work linked, the last messages of its coding threads, older messages in the conversation, and its notes and memory. It shows you only work from the conversation you are in, work you asked for, and, if your link may read GOAT’s threads, the rest.
With every message GOAT Bot also sees a short journal of the latest work you may see: what was asked, how it ended, and its pull requests.
Ask GOAT Bot “what’s pending?” to see everything still open: queued, running, waiting on you, or finished but not yet reported to you. Say “stop everything” (or name one piece of work) to stop it, including anything it handed to subagents, on this machine or another one. A machine still on an older GOAT stops only the current turn; subagents it started there finish on their own.
When more than one machine is connected, GOAT Bot spreads research across the ones with room to spare and tells you which machine ran what. It never starts a Boat cloud machine unless you ask.
GOAT Bot runs up to four pieces of your work at once. More wait in order and start as earlier ones finish; the reply says when something is queued.
Settings → GOAT Bot → Automatically choose a worker computer lets work use an already-connected SSH computer with the same repository and requested model. It is off by default. Turning it on allows the worker’s instructions, notes and named skills to go to the selected computer. GOAT prefers ready accounts with more quota left and less channel work, and shows why it chose one. Naming a computer keeps work there. Unknown or unavailable repositories are skipped; GOAT does not prepare computers or sign in for this choice. Retrying automatically placed work chooses again; a running worker never moves automatically.
Work counts as waiting on you when the agent working on it, or a subagent it started in GOAT, needs your approval or an answer. That includes a subagent running in GOAT on another machine. Subagents that run on another service are not counted.
When work needs an approval, GOAT Bot sends a card saying what the agent wants to do. Tap Approve or Deny in GOAT. In Slack, tap the button or reply with its number (2 is the second choice), or reply “approve” or “deny”. The same number works in iMessage, which has no buttons. If several approvals are waiting, add the card’s code, such as “approve A2”. Only people with the same access as whoever asked for the work can answer. An approval on another GOAT machine is sent there; if that machine can’t take it, answer it in that thread.
Ask GOAT Bot to remember a standing rule (“always reply in Korean”) and it follows it from then on, in Slack and in GOAT. It keeps a note only when the message you send asks for it, and each note names the message it came from. Notes live in goat-bot/memory.md in GOAT’s data folder on that machine; edit the file to change them. Put written procedures in goat-bot/skills/ as Markdown files, and the bot points its coding agents to the one that fits. Work sent to another machine gets the notes, and any procedure the request names, in its first message.
If a machine running your work goes to sleep or disconnects, GOAT Bot tells you. Say “continue on ” to continue from the last copy of its work, taken each time the work finishes a turn. New files over 2 MiB and new files that look like secrets are not copied. The work on the original machine is not deleted.
When a provider has more than one signed-in account on the machine running the work, GOAT uses the one with more quota left and the task card says why, such as “Codex (work) · 20% weekly used”.
For memories used in coding conversations and shared between devices, see Memory.
Work you start from Slack is capped at the access level chosen when you created the link code, or at Access without a link if you have not linked. Its coding agents can use GOAT memory only as far as the link allows. A link you narrow later, or switching to Linked accounts only, applies to work that has not started yet.
When you connect to several machines, Slack runs only on the machine GOAT Bot lives on. Change it in Settings → Integrations → GOAT Bot; set up Slack on that machine.
If the machine GOAT Bot lives on is gone or stays disconnected, GOAT Bot stops answering on every machine and its screen says so. Use Move GOAT Bot to … there, or pick another machine in Settings → Integrations → GOAT Bot. An agent with full access can do the same on the machine it runs on with goat_call {path:"channel.bot_home_claim"}. Other machines follow the move once a GOAT app connected to them is open.
When GOAT Bot can’t read a message, its reply says why, such as a model that needs you to sign in again or a usage limit, so you know whether resending helps.

iMessage (Mac)

GOAT can read direct text messages from allowed iMessage handles on the Mac running its server and reply in the same conversation. Group chats and attachments are not supported. An approval or question lists choices as 1., 2., and so on; reply with that number to choose it. Give the app that launches the server Full Disk Access to read Messages. macOS asks separately for Automation access when GOAT sends through Messages; Messages must be signed in.

Run goat imessage enable --allow +15551234567 on that Mac, then goat imessage link-code. From the allowed phone or email, send link 123456 to the Mac’s Messages account, replacing the example with the printed code. The connection starts with an approval-required access ceiling; --full-access on enable is an explicit opt-in. Restart an already running server after enabling. goat imessage status shows whether the channel is enabled and whether the local database can be opened without printing handles or message text. Run goat imessage disable to stop it and goat imessage unlink <handle> to revoke a linked sender. An installed server uses its own data directory; pass --base-dir when managing a different local server.