SlackAgentBridge

Slack Agent Bridge

Control local [Claude Code](https://claude.com/claude-code), [Codex CLI](https://developers.openai.com/codex/cli/), and [Pi](https://github.com/earendil-works/pi) sessions from Slack. Each native session gets a private Slack channel where prompts, responses, progress, and attachments flow both ways.

Version 2 has one command language everywhere: `sab` in a shell and `/sab-*` in Slack. The active Slack channel selects its provider; only creation and provider switching need an explicit `claude`, `codex`, or `pi` target.

Provider processes live in detached-capable tmux sessions. Ghostty is an optional viewport, not a process-lifetime requirement: close every terminal and the agents continue running; open or focus one later without resuming or duplicating the native conversation.

[!WARNING] **This is remote code execution by design.** Slack-spawned Claude sessions default to `--dangerously-skip-permissions`; Slack-spawned Codex sessions default to `--dangerously-bypass-approvals-and-sandbox` (`--yolo`). Pi's built-in tools are unrestricted by default; SAB's optional `--safe` flag adds fail-closed Slack approval per tool call. Anyone able to act as the bridge owner can steer processes with that Mac user's privileges. Read [SECURITY.md](SECURITY.md) before installing.

[!NOTE] The daemon currently targets macOS and launchd. Ghostty is needed only when terminal viewports are wanted. Claude uses its Channels API; Codex uses hooks, tmux, and a loopback App Server event proxy; Pi uses an explicitly loaded native extension.

Capabilities

Capability	Claude Code	Codex CLI	Pi
Private channel per native session	✓	✓	✓
Slack prompts and attachments	✓	✓	✓
Return generated files to Slack	✓	✓	✓
Final responses and live working status	✓	✓	✓
Selected interim progress	✓	✓	managed runs
Model and effort controls	✓	✓	✓
Remote permission decisions