Future Ideas

Future Ideas

Things we want to build, contribute upstream, or explore.

Contribute Multi-Agent Recipes to OpenClaw

The current OpenClaw multi-agent docs explain the mechanics well (bindings, sandbox, elevated, etc.), but don’t provide end-to-end recipes for common real-world patterns. Users have to piece together config from multiple doc pages:

• concepts/multi-agent.md — routing basics
• multi-agent-sandbox-tools.md — per-agent restrictions
• tools/elevated.md — elevated allowlists and defaults
• gateway/configuration.md — schema reference

This makes it hard to set up something like “a family assistant where I have full access but my kids are sandboxed.”

Proposed: Recipe Section in OpenClaw Docs

Add a Recipes or Patterns section to the OpenClaw docs with complete, copy-paste-ready examples.

Example Recipe: Family Assistant

Goal: One assistant identity, different permission levels:

• Owner gets full access everywhere
• Family members get sandboxed (no exec, no file writes, no sensitive tools)
• Group chats sandboxed, but owner can @mention for elevated access

Architecture:

	main-agent	group-agent	family-agent
Routes	Owner DMs, Webchat	All group chats	Family member DMs
Sandbox	Off	All	All
Tools	Full access	Restricted + elevated for owner	Restricted, can escalate to main

Key Config Snippets:

{
  "agents": {
    "defaults": {
      "elevatedDefault": "on"  // Owner auto-elevated when on allowlist
    },
    "list": [
      {
        "id": "main",
        "sandbox": { "mode": "off" }  // Owner DMs unsandboxed
      },
      {
        "id": "groups",
        "sandbox": { "mode": "all", "scope": "agent" },
        "tools": {
          "profile": "minimal",
          "alsoAllow": ["web_search", "web_fetch", "tts"],
          "deny": ["exec", "write", "edit", "gateway"]
        }
      },
      {
        "id": "family",
        "sandbox": { "mode": "all", "scope": "agent" },
        "subagents": { "allowAgents": ["main"] },  // Can escalate
        "tools": { /* similar restrictions */ }
      }
    ]
  },
  "tools": {
    "elevated": {
      "enabled": true,
      "allowFrom": {
        "bluebubbles": ["+1XXXXXXXXXX"]  // Owner's phone number
      }
    }
  },
  "bindings": [
    { "agentId": "family", "match": { "channel": "bluebubbles", "peer": { "kind": "dm", "id": "+1XXXXXXXXXX" }}},
    { "agentId": "groups", "match": { "channel": "bluebubbles", "peer": { "kind": "group", "id": "*" }}},
    { "agentId": "main", "match": { "channel": "bluebubbles" }}  // Catch-all for owner
  ]
}

Gotchas We Discovered:

1. tools.elevated.allowFrom is a record by channel, not an array (e.g., { "bluebubbles": ["+1234567890"] })
2. Group wildcard binding requires "id": "*" (not just "kind": "group")
3. elevatedDefault lives in agents.defaults, not per-agent
4. Binding order matters — specific matches before catch-all

Other Recipe Ideas

• Work/Personal Split — separate workspaces, different calendar access
• Public Bot — fully sandboxed, no file access, rate limited
• Dev/Prod Agents — same config, different auth profiles

WhatsApp as a Full Channel

Currently WhatsApp monitoring is read-only with reactions. Future work:

• Full send/receive through WhatsApp (not just monitoring)
• Multi-channel routing (iMessage + WhatsApp to the same agent)
• Per-channel response formatting (WhatsApp supports markdown, iMessage does not)

Proactive Intelligence

• Morning briefing cron job (weather, calendar, travel, packages)
• Smart email follow-up detection (“you said you’d respond to X”)
• Expense categorization from email receipts to YNAB
• Flight disruption monitoring with automatic rebooking suggestions