| name | personal-assistant-stack |
| description | How Walter-OS thinks about a personal AI assistant that handles inbound messages (WhatsApp/Telegram/iMessage/Instagram), can reply for you, surfaces the urgent stuff, creates calendar events from chat context, and uses your LLM subscription rather than per-token API spend. Use this skill when the user asks "what's my personal assistant strategy", "should I use OpenClaw vs a custom bot", "how do I let the AI reply to my messages", "how do I route AI cost via subscription". Decision tree below. |
Walter-OS personal assistant stack — the answer is one assistant, not three
The core decision
Three Telegram bots is the wrong shape. The right shape:
| Bot | Role | Surface |
|---|
@${WALTER_TELEGRAM_BOT_HANDLE} | Outbound alerts only | Walter-VM monitors → push to me |
@${WALTER_OPENCLAW_BOT_HANDLE} | Personal assistant — bidirectional | I talk to it, it reads my messages, it replies for me |
No third bot. OpenClaw IS the personal assistant. Adding a third loses the
context boundary (the assistant can't see what the alerter sent without a
relay) and adds another auth surface to manage.
What OpenClaw covers
OpenClaw (openclaw.ai, self-hosted, Node 24+) connects to ALL of:
- WhatsApp (via Beeper or direct WA Business API)
- Telegram (the bot for outbound, plus your user account for read-only inbox)
- Slack, Discord, Google Chat
- iMessage, Signal (via Beeper bridge)
- Microsoft Teams, Matrix, Mattermost
- Instagram DMs (via Beeper bridge)
It's bidirectional: you message it, it reads on your behalf, it replies
when you authorize, it summarizes when you ask, it pings you on what matters.
How it answers your specific use cases
"Reply to messages for me when I'm busy"
Pattern: pairing + pre-approved auto-reply rules.
- New WhatsApp arrives → OpenClaw scores priority (urgent / can-wait / ignore)
- For "can-wait": auto-reply "Hey, I got your message and will answer in X hours"
using a template you defined for the contact's locale and context.
- For "urgent": ping you on Telegram with summary + suggested reply, you
approve via inline button → assistant sends.
- For "ignore" (newsletters, chain forwards): silent, queued for end-of-day
digest.
The pattern is never autonomous send without approval for new contacts /
unusual context. That's the safety bar.
"Tell me what's important"
Daily digest format (you can ask anytime, or schedule):
Daily — May 4, 2026
🔥 Urgent (action needed today)
• Maria (WhatsApp): asks for the public-tender proposal by Friday 6pm
• Marcos (Slack #[company]-eng): Geyser plugin crashed in mainnet, needs review
📬 Worth seeing (no action)
• Telegram: 3 group messages in #buenos-aires-tech
• Email: Stripe invoice PDF (auto-archived)
• Discord: tagged in #anaconda-team about MVP review
🗓️ Scheduled today
• 16:00 — call with potential investor (Alex C.)
• 18:00 — gym
🌚 Quiet
• 47 messages auto-handled (low-priority replies, newsletters, etc.)
"Schedule a coffee at 5pm Thursday from my chat with X"
OpenClaw + Calendar API integration:
- You: "OpenClaw, schedule coffee with Maria Thursday at 5pm"
- OpenClaw extracts: title="Coffee with Maria", when="next Thursday 17:00 BA"
- Confirms: "I will add 'Coffee with Maria' to your Personal calendar for Thursday, May 8 at 17:00. OK?"
- You: 👍
- Inserts event via Google Calendar API.
Or fully passive: OpenClaw watches your Telegram chat with Maria, sees
"sure, Thursday at 5pm at Lattente?" → drafts the event → asks you to approve.
The calendar integration uses a Google Service Account scoped to a single
calendar (Personal), not your full account. Limits blast radius.
Beeper — the unified-message layer
Beeper is NOT the assistant — it's the transport. Beeper bridges
WhatsApp/iMessage/Instagram/Signal/etc. into a single Matrix-compatible
inbox. OpenClaw connects to Beeper's local Matrix endpoint (instead of
each individual service) and gets one unified stream.
Why this matters:
- WhatsApp alone has no public API for personal accounts. Beeper bridges it
legitimately via your phone session.
- iMessage requires a Mac running Beeper Mini relay or similar.
- Instagram doesn't expose DMs to apps directly. Beeper bridges it.
So the actual stack is:
┌──────────────────────────────────────────┐
│ Phone (your real WA/iMessage/IG sessions)│
└──────────────────┬───────────────────────┘
│ bridges / relays
▼
┌──────────────────────────────────────────┐
│ Beeper Cloud (Matrix homeserver) │
│ https://beeper.com — paid $10/mo OR │
│ self-hosted (more work, free) │
└──────────────────┬───────────────────────┘
│ Matrix protocol
▼
┌──────────────────────────────────────────┐
│ OpenClaw on Walter-VM │
│ - reads incoming Matrix events │
│ - LLM via LiteLLM (Sonnet by default) │
│ - actions: reply, summarize, tag, queue │
└──────────────────┬───────────────────────┘
│
┌───────┴────────┐
│ │
@${WALTER_OPENCLAW_BOT_HANDLE} Google Calendar API
(your inbox to it) (event creation)
Beeper Mini (the cheaper iOS app) was killed by Apple. Beeper Cloud
($10/mo) is the only viable path for iMessage. WhatsApp/IG bridges work
on either tier.
Where the cost lives
LLM cost — use Sonnet by default, Opus only when reasoning matters
Critical principle: don't burn your Opus quota on routine assistant tasks.
| Task | Model | Why |
|---|
| Auto-reply scoring (urgent/can-wait/ignore) | Sonnet | Pattern recognition, no deep reasoning |
| Summarize a chat | Sonnet | Compression, no deep reasoning |
| Draft a reply in casual localized style | Sonnet | Style, no deep reasoning |
| Calendar event extraction | Sonnet | Structured extraction |
| Write a long-form blog draft | Opus | Long context, quality matters |
| Plan complex multi-step task | Opus / Opus-thinking | Reasoning |
| Code review across 3 files | Opus | Reasoning + context |
LiteLLM virtual keys per use case, each pinned to a specific model:
virtual_keys:
- key: openclaw-routine
models: [claude-3-7-sonnet]
budget_max: 30/mo
- key: openclaw-reasoning
models: [claude-opus-4]
budget_max: 5/mo
- key: operator-cli-personal
models: [*]
budget_max: 50/mo
- key: n8n-workflow-runner
models: [claude-3-7-sonnet]
budget_max: 20/mo
Total budget: $105/mo cap across all automation. You see the dashboard at
llm.${WALTER_DOMAIN}, can yank any key if it runs hot.
Subscription proxy — the honest answer
Anthropic does NOT support subscription-as-API. Claude Pro / Max
subscriptions cover Claude.ai web + Claude Code OAuth + Claude Desktop
OAuth. They are NOT exposed as a generic OpenAI-compatible endpoint.
Workarounds exist (claude-code-router, anthropic-cookie-relay, etc.) but:
- They scrape session cookies from a logged-in Claude Code instance.
- They violate Anthropic ToS (subscriptions are explicitly per-user, not for
programmatic 24/7 use).
- Anthropic actively rate-limits and bans accounts caught doing this.
Your subscription stays for interactive personal use (Claude Desktop,
Claude Code on your laptop). For automation (n8n, OpenClaw, scripts), use
the Anthropic API key with budget caps in LiteLLM. Total monthly bill:
~$15-50/mo depending on volume. Cheaper than paying $200/mo for Claude
Max if your automation isn't constant.
If you really want to keep subscription-only:
- Run OpenClaw on Walter-VM with Claude Code OAuth: technically possible
(OpenClaw supports OAuth providers), but the OAuth session lives in
Claude Code's sandbox and isn't a stable transport for a 24/7 daemon.
- ChatGPT subscription via OpenClaw: explicitly supported per OpenClaw
README. You log in to ChatGPT once, OpenClaw reuses that session. Same
caveats around ToS.
Operator's pragmatic choice: keep Anthropic API + LiteLLM caps for
automation. Use Claude Pro for personal CLI / Desktop work. They serve
different threat models.
Trust + safety boundaries
| Action | Default | Override |
|---|
| Read your messages for summary | ✅ allowed (operator-explicit invocation) | n/a |
| Reply with auto-template | ⚠️ requires per-contact opt-in (OpenClaw, allow auto-reply on this thread) | n/a |
| Send a new message you didn't draft | ❌ never | one-shot inline button approval |
| Create a calendar event | ✅ with confirmation | always |
| Modify an existing event | ⚠️ confirm | always |
| Pay for something | ❌ never | not even with confirmation — out of scope |
| Log into a new account | ❌ never | not even with confirmation |
| Read messages in chats with sensitive third parties ([Project B] patients, etc.) | ❌ blocked | manual override + logged |
Keep this list short. The longer it gets, the more Stockholm-syndrome you get
into auto-overriding. Tight default = good asistente.
Phase K4 deployment plan
1. Beeper Cloud account ($10/mo)
- Add WhatsApp bridge (scan QR with phone)
- Add iMessage bridge (Beeper Mac app installed)
- Add Instagram bridge (login)
- Add Telegram bridge (your user account, optional)
2. OpenClaw on Walter-VM
setup/walter-host/services/openclaw/
compose.yml - Node 24+ container
env.template - LITELLM_BASE_URL, BEEPER_HOMESERVER_URL,
BEEPER_TOKEN, ANTHROPIC_API_KEY (via LiteLLM)
deploy.sh - idempotent
URL: claw.${WALTER_DOMAIN} (CF Tunnel + Access)
3. LiteLLM virtual keys
- Create openclaw-routine (Sonnet, $30/mo cap)
- Create openclaw-reasoning (Opus, $5/mo cap)
- Inject via Infisical to OpenClaw env
4. Pairing
- Operator messages @${WALTER_OPENCLAW_BOT_HANDLE} from Telegram
→ bot replies with pairing code, links Telegram identity to OpenClaw
- Same flow for Beeper bridge: OpenClaw asks "I see WhatsApp messages,
authorize?" via Telegram bot; you approve once.
5. Calendar integration
- Create Google Service Account in walter-os GCP project
- Share "Personal" calendar with the SA's email
- Save SA JSON to Infisical workspace=walter-os env=dev key=GOOGLE_SA_JSON
- OpenClaw reads it for calendar.events.insert calls
6. Daily digest schedule
- Cron in OpenClaw: 09:00 BA → assemble overnight digest → send to bot
- Cron 22:00 BA → end-of-day summary
ETA when Walter-VM is ready: ~3-4 hours of operator time + ~2 hours of agent
time (configs + scripts).
Failure modes to expect
- Beeper bridges break occasionally (WhatsApp pushes new TLS pinning,
iMessage Apple updates). Plan for ~1 outage/quarter.
- LLM hallucinates priority scoring for messages from unfamiliar
contacts. Tune the system prompt + add "low confidence → ask operator".
- Time zone confusion — calendar event extraction with "next Thursday"
when it's Sunday in Argentina but Saturday in operator's location after
travel. Always confirm timezone in event description.
- Group chats — auto-reply in groups is socially weird. Default: mute
auto-reply for groups, only DMs.
What this skill does NOT cover
- Voice / phone calls — out of scope. Use call screening apps if you need it.
- Email auto-reply — Gmail has its own; OpenClaw can read for digest, but
reply-by-email is high-risk (one wrong "Reply All" can fire ruined careers).
- Meeting transcription — separate tool (Otter, Granola, Fathom).
References
