If you search “how to use OpenClaw for small business operations automation,” you probably have the same headache I did: too many apps, not enough hours, and no budget for another hire. I spent two weeks wiring OpenClaw into the tools that actually run my consultancy. It now handles 80 % of the chores a virtual assistant used to do. Below is the exact playbook—commands, configs, rough costs, and what broke along the way—so you can skip the trial-and-error phase.
Why bother? ROI vs hiring an admin assistant
The median US salary for an administrative assistant is USD 45 K/yr, before benefits and churn costs. My OpenClaw bill on ClawCloud last month was USD 97.43. Even after adding paid APIs (SendGrid, QuickBooks, Calendly) the fully loaded cost stays under USD 200/mo. The breakeven point for me was the first week I didn’t forget to send an invoice.
- Fewer context switches: one chat thread instead of six dashboards.
- 24×7 availability: the bot answers client pings at 02:00 without “overtime”.
- No single point of human failure: vacations and sick days are not project risks now.
Downside: setup takes focus, and you will expose API keys you absolutely must lock down. More on security later.
Fast setup: OpenClaw running in 10 minutes
1. Sign up on ClawCloud
The hosted version saves you from dealing with Node 22 upgrades and TLS certificates. Name your agent, pick the small-biz-pro template, and ClawCloud provisions the gateway + daemon in about 60 seconds.
2. Local fallback (optional)
If you prefer on-prem, grab the binary:
$ nvm install 22
$ npm i -g openclaw@2.8.3 # latest as of 2024-06-01
$ openclaw init --template small-biz-pro
Both paths generate an .env with placeholder secrets for Gmail, QuickBooks, Calendly, and Slack. Do not commit this file. Add it to .gitignore immediately.
Connecting the core apps
OpenClaw leans on Composio’s 800 integrations. Below are the ones that moved the needle for me, plus caveats.
Email (Gmail)
$ openclaw connect gmail
? Paste OAuth client ID: ****
? Paste client secret: ****
? Scopes: readonly, send
Readonly + send scopes are enough for triage and outbound replies. Anything higher triggers extra reviews during Google’s OAuth audit—avoid.
Calendar (Google Calendar)
$ openclaw connect gcal --scopes events.readonly,events.write
Stick to one calendar to simplify scheduling. Multiple calendars work but timezone math inside OpenClaw 2.8.3 is flaky if names differ only by case; bug filed on GitHub #8127.
Invoicing (QuickBooks Online)
$ openclaw connect quickbooks --realm 1234567890
QuickBooks’ OAuth tokens expire every 60 minutes. The daemon refreshes them automatically, but you must set QB_REFRESH_TOKEN in .env.
Chat (Slack)
I’m running all agent comms from a dedicated Slack channel #backoffice-bot. OpenClaw posts status, asks for clarifications, and dumps daily digests there.
Building the workflows that actually save time
You could let GPT-4 “figure it out,” but vague prompts create vague results. Below are the concrete tasks I automated and the YAML I shipped.
Email triage + tagging
name: inbox_zero
on:
schedule: "*/15 * * * *" # every 15 min
run:
- gmail.list(query: "is:unread -category:promotions -label:bot-skipped")
- foreach email in $results:
if email.from in [vip@acme.com, accountant@example.com]:
gmail.modify(email.id, addLabels: ["VIP"], removeLabels: ["INBOX"])
else:
openclaw.llm("Summarize and draft a reply if needed")
gmail.modify(email.id, addLabels: ["REVIEW"])
The LLM step uses cl100k tokenizer limits—stay under 12 k tokens or Google complains. Summaries and drafted replies land in Slack for a quick eyeball before sending.
Automated invoicing on project milestones
name: invoice_on_commit
on:
github: push
filters:
branches: [main]
paths: ["/milestones/**"]
run:
- quickbooks.createInvoice(customer: "{{ commit.author.email }}",
amount: calculateAmountFromYaml(commit.diff),
due: "NET15")
- slack.post(channel: "#backoffice-bot", text: "Invoice drafted for {{ commit.author.email }} → ${{ amount }}")
Yes, a Git commit can trigger an invoice. Clients love the instant transparency; accountants love the audit trail.
Meeting scheduling without the ping-pong
name: auto_schedule
on:
email: received
filters:
subject: /.*(call|meeting|sync).*/i
run:
- openclaw.llm("Extract requested times, preferences,
and locations from the email → JSON")
- gcal.findSlot(duration: 30, within: "next 10 days", workingHours: "09:00-17:00")
- calendly.createInvitee(name: email.sender, slot: $slot)
- gmail.send(to: email.sender, template: "schedule_confirm", slot: $slot)
The built-in NLP handles natural-language time zones surprisingly well, but it still misreads “next Friday” around midnight UTC. I force clients to pick from a Calendly link in the confirmation email to avoid accidental double bookings.
Customer communication across channels
One of OpenClaw’s party tricks is bridging WhatsApp ↔ Telegram ↔ Slack. For a retailer running customer support this alone can axe a third-party SaaS line item.
WhatsApp + Telegram unified support
name: omni_support
on:
whatsapp: message
telegram: message
run:
- memory.getOrCreate("customerProfile", key: message.sender)
- openclaw.llm("Generate helpful response using profile and FAQ.md")
- reply.to(message)
- if message.sentiment == "frustrated":
escalate.to(slack, channel: "#support-escalations")
Sentiment scoring uses OpenAI text-embedding-3-small. Cheap and good enough. Escalations keep the human loop where it matters.
Social media autopilot (but keep your voice)
Twitter drafts from ChatGPT still read like ChatGPT. I solved it by feeding the bot 100 previous posts and instructing it to mirror my sentence length, sarcasm level, and Oxford-comma stance.
name: tweet_queue
on:
schedule: "0 13 * * 1-5" # lunch hour weekdays
run:
- notion.query(database: "ContentIdeas", filter: {Status: "Ready"})
- foreach idea in $results:
openclaw.llm("Write a 280-char post in my tone. Avoid #growthhack.")
twitter.post(content: $output)
notion.update(idea.id, {Status: "Posted"})
The Twitter API v2 is now pay-walled. I pay USD 100/mo for elevated write access—a cost worth absorbing since one viral tweet brought in a five-figure contract.
Bookkeeping & reporting without spreadsheets
Every Friday at 16:00 the bot ships a single PDF to my inbox: P&L YTD, overdue invoices, upcoming tax liabilities. Source of truth stays in QuickBooks; OpenClaw just queries and formats.
name: weekly_finance_pack
on:
schedule: "0 16 * * FRI"
run:
- quickbooks.report(type: "P&L", period: "YTD")
- quickbooks.report(type: "OpenInvoices")
- openclaw.llm("Generate executive summary under 200 words")
- pdf.render(reports + summary)
- gmail.send(to: "me@example.com", subject: "Weekly finance pack", attach: pdf)
No more Monday-morning surprises. If cash flow dips below a threshold, the bot pings Slack with :@alert:.
Security, governance, and not getting fired by your accountant
- Principle of least privilege: each integration has its own OAuth credentials with minimal scopes. Don’t reuse secrets across environments.
- Audit trails: enable
OPENCLAW_LOG_LEVEL=infoand ship JSON logs to Datadog. Auditors like immutable evidence. - Human approval gate: for any action above USD 500 or with external recipients, I route through
slack.approvals. Two clicks, done. - Data residency: ClawCloud lets you pin your workspace to
eu-west-3. Do it if GDPR keeps you up at night.
Real-world testimonial: “the bot runs my company”
Chris Lysne (tiny SaaS founder, solo) posted in GitHub Discussions last month: “I merged the gateway with my Go microservice and now ship features instead of formatting invoices—OpenClaw literally runs my company.” His metrics:
- 14 hrs/week saved on admin
- USD 860/mo SaaS stack trimmed (Zapier, Intercom, Buffer)
- First non-founder hire pushed back by six months
The point: this isn’t theory. Small teams are already operating like 10-person shops because their bot refuses to sleep.
Cost breakdown & breakeven calculator
- ClawCloud base plan: USD 49/mo
- Extra GPU quota (LLM 8 K context): USD 29/mo
- OpenAI tokens (gpt-4o 128k): ~USD 12/mo for 400k tokens
- Twitter API v2 elevated: USD 100/mo (optional)
- QuickBooks Simple Start: USD 30/mo
Total: USD 220/mo. Compare that to a part-time assistant at USD 1 500–2 000/mo. Your breakeven is ~1.5 weeks if you value your own time at USD 50/hr and reclaim just 1 hour/day.
Next step: start with one pain point, not seven
Pick the task you hate most—email triage, invoice chasing, appointment ping-pong—and wire that first. Once it survives a week in production, layer on the next workflow. OpenClaw scales down as gracefully as it scales up, so there’s no penalty for starting small. Your future self (and your P&L) will thank you.