Business Agent Launchpad#

Use this skill when the user wants a first-run business setup for OpenClaw, Hermes, or a similar agent harness.

The goal is to translate ordinary business language into:

• scenario and workflow description
• agent roles and operating rules
• object-based Markdown memory
• tool and permission plan
• model tier and escalation plan
• free vs paid cost choices
• OpenClaw/Hermes starter artifacts

First Check#

Verify the CLI exists:

bash

agent-launchpad doctor

If it is not installed, find the project folder and install it:

bash

npm install -g .

Then re-run:

bash

agent-launchpad doctor

Recommended Flow#

For a non-technical user, run the interview:

bash

agent-launchpad interview --out ./first-business-agent

For a quick preset:

bash

agent-launchpad generate --preset sales --tier free --runtime both --out ./sales-agent

Available scenario keys:

• sales
• admin
• finance
• legal
• project
• hr
• customer_service
• executive_assistant
• procurement

Available tiers:

• free: Markdown memory and local tools first
• local_plus: local semantic memory with a small embedding model
• cloud_plus: paid models, hosted embeddings, and stronger document understanding
• team: shared memory, permissions, audit, and business-system integration

How To Choose#

Use free when the user is just starting, wants transparency, or lacks API keys.

Use local_plus when the user has many documents and cares about privacy.

Use cloud_plus when document quality, long-context reasoning, OCR, or lower setup friction matters more than API cost.

Use team when multiple people share memory or the workflow touches customer records, HR, finance, contracts, CRM, ERP, or audit requirements.

After Generation#

Open the generated README.md, then inspect:

• route-map.md
• cost-and-model-plan.md
• permission-plan.md
• tool-plan.md
• memory/index.md
• openclaw/ or hermes/

Verify:

bash

agent-launchpad verify ./first-business-agent

Safety Rules#

• Do not let the generated agent send messages, submit forms, delete files, change CRM/accounting records, approve payments, or store sensitive data without explicit user approval.
• Keep Markdown object memory as the human-readable source of truth.
• For paid models, hosted embeddings, cloud OCR, or managed vector databases, explain the benefit and tradeoff before enabling them.
• Treat legal, finance, HR, customer-impacting, and executive work as review-support workflows, not autonomous decision workflows.