agent-slack

agent-slack

Use agent-slack from $PATH. If it is missing, install it with:

curl -fsSL https://raw.githubusercontent.com/stablyai/agent-slack/main/install.sh | sh

Fallback: npm i -g agent-slack (Node >= 22.5).

Run agent-slack --help or the relevant subcommand help before guessing a command or flag. If a capability named here is absent from installed help, report version skew instead of guessing. Do not self-update the CLI without explicit authorization.

Safety

• Read and search freely.
• Perform write actions only when explicitly requested: sends, edits, deletes, reactions, invitations, channel or canvas creation/editing, mark-read operations, scheduling or canceling delivery, uploads, Later state/reminder changes, DM/group-DM creation, and workflow run. Workflow runs can execute downstream actions.
• For compose- or review-only requests, return proposed text without invoking Slack, or use message draft create to add a Slack-native draft the user can review and send (nothing is posted). message compose is send-capable; use it only when the user explicitly asks to open the interactive editor. In CI or another noninteractive environment, do not invoke it without separate authorization to send immediately: CI skips the editor and sends supplied text.
• With AGENT_SLACK_SAFE_MODE=1 (or the global --safe-mode flag) set, safe mode is enforced at the tool level: message send is redirected to the draft editor and message edit/message delete are blocked. Use it when nothing should post without human review.

Workflow

1. Run agent-slack auth whoami. If needed, import credentials with auth import-desktop, auth import-brave, auth import-chrome, or auth import-firefox, then run auth test.
2. Prefer a Slack message URL when one is available. It carries the workspace, channel, and timestamp needed by most message operations.
3. Choose the narrowest read operation: message get for one message, message list for a full thread or channel history, and search messages or search files for discovery.
4. Use output limits such as --limit, --max-body-chars, and --max-content-chars to avoid unnecessary context.
5. For a requested write, execute only the requested mutation and verify the resulting JSON metadata.

For scheduled writes, prefer --schedule with an ISO 8601 timestamp and explicit offset when timezone matters. Named --schedule-in phrases use the executing environment's local timezone; confirm that it matches the user's intent.

Named later remind --in values such as tomorrow or monday also use the executing environment's local timezone at 9:00. Confirm that timezone or pass an explicit Unix timestamp.

Use --no-unfurl with message send or message compose when the user wants Slack link and media previews suppressed. It cannot be combined with message send --attach.

Ordinary message send and message edit calls auto-convert lists. message send --blocks and message edit --blocks use supplied Block Kit blocks, while message send --attach sends its initial comment without automatic list conversion. Inside auto-converted lists, use Slack's <URL|label> syntax because CommonMark [label](URL) links are not converted into labeled link elements.

Slack-native drafts (message draft list|create|update|delete) manage drafts that appear in the user's Slack client; create posts nothing. create and update accept repeatable --attach <path>; on update the files are added to the draft's existing attachments rather than replacing them. They use undocumented session endpoints and require browser-style auth (xoxc/xoxd).

canvas edit uses Slack's public canvases.edit API and applies exactly one operation per call. The default replace operation replaces the whole canvas; section-targeted inserts/replacements and deletes require the section ID returned by Slack's Canvas tooling, while rename takes --title. Content operations take exactly one --file or --markdown source. It requires a standard token with canvases:write; imported browser credentials can create standalone canvases but cannot edit.

Conditional references

• Read references/targets.md only when choosing between a message URL, channel, or user target, or when resolving multiple workspaces.
• Read references/output.md only when handling returned message or canvas metadata, resolved users, or downloaded and failed attachments.