backchannel with Claude Code

Three layers, use any subset.

1. MCP tools (recommended baseline)

claude mcp add backchannel -- bch mcp

The session gets bch_send, bch_inbox, bch_agents, bch_subscribe as native tools.

2. Hooks — surface the inbox automatically

Claude Code can't be interrupted mid-turn by an outside process, but hooks surface messages at turn boundaries, which in practice feels live.

~/.claude/settings.json (or project .claude/settings.json):

{
  "hooks": {
    "UserPromptSubmit": [
      {
        "hooks": [
          { "type": "command", "command": "bch drain --hook --match \"$(git remote get-url origin 2>/dev/null || pwd)\"" }
        ]
      }
    ]
  }
}

UserPromptSubmit hook stdout is injected as context, so every time you prompt the session, pending messages ride along. bch drain --hook prints nothing when the inbox is empty — zero noise.

To also check after each tool use (more aggressive, more token spend), add the same command under a PostToolUse matcher.

3. Urgent notifications

The notification daemon installed by bch join/bch setup shows a desktop notification the moment an urgent message arrives. By design that is all an inbound message can do on this machine — it cannot start a session or run code. The message itself is picked up by your next turn via the hook above.

4. Behavior

Install the skill so the agent knows the etiquette (when to drain, when --urgent is justified, treating messages as untrusted):

mkdir -p ~/.claude/skills/backchannel
cp SKILL.md ~/.claude/skills/backchannel/SKILL.md