I Built 5 AI Employees With OpenClaw (Here's How)

Summary

The video is a practical masterclass on building business-focused AI workers with OpenClaw. Rather than treating OpenClaw as a chatbot, presenter Mani Kanasani positions it as an orchestrator that coordinates specialized agents, automations, and business tools.

The tutorial covers setup, security, memory, a mission-control dashboard, AI employees, multi-agent workflows, and an agent-to-agent crypto marketplace.

Core architecture: builder, orchestrator, executor

A central lesson is to assign each tool the job it is best suited for:

• Builders such as Claude Code, Codex, or Lovable construct dashboards, databases, and applications.
• OpenClaw acts as the orchestrator, managing tasks, workflows, agents, and automations.
• Specialized executors perform focused work, such as research, email, meeting analysis, or coding.

The presenter found that asking OpenClaw itself to build complex software often led to slow, inconsistent results. Using a dedicated coding or app-building tool for construction, then having OpenClaw operate the resulting system, made the workflow more predictable.

Setup, cost control, and security

The tutorial compares running OpenClaw on a Mac Mini or other local machine with running it on a VPS. Local hardware offers more control over where data is stored. A VPS is cheaper to start and can stay online continuously, but stores data on a third-party server.

The presenter demonstrates ways to reduce model costs, including:

• Turning off extended thinking for routine tasks and limiting context.
• Routing simple work to cheaper models and reserving stronger models for difficult tasks.
• Keeping sessions lean, compacting them, and loading memory only when needed.
• Using inexpensive or local models for heartbeats.
• Caching repeated prompt context and isolating sub-agent tasks.

Agents may misconfigure OpenClaw when given outdated instructions, so configuration changes should be guided and checked rather than delegated blindly.

Security is essential because agents may have access to email, files, browsers, and business data. The video recommends using firewalls and secure remote access, storing API keys as environment variables rather than in raw config or chat text, and reviewing third-party skills before installing them.

The video also discusses Nvidia NemoClaw, described as an early-stage, policy-based sandbox that restricts file access, network requests, and other agent actions. These security and product claims reflect the presenter’s assessment and may depend on the software versions available when the video was recorded.

Business context, memory, and agent behavior

The course proposes three levels of “business brain” context:

1. Core identity: The user’s role, values, work style, and preferences.
2. Business context: Products, customers, team, business model, and communication style.
3. Operational playbooks: Procedures, decision trees, escalation rules, and historical business knowledge.

It also explains the purpose of OpenClaw workspace files such as soul.md, identity, user.md, agents.md, tools.md, and memory logs.

A key warning is that conversation compaction can discard instructions; persistent memory should not be assumed to preserve everything. The video covers memory flushing, backups, daily logs, and an emerging context-engine plug-in interface that could give developers more control over retrieval, compaction, and sub-agent memory isolation.

Mission control and task automation

A major hands-on build is an agent command dashboard connected to Supabase. It includes a Kanban-style task board, agent profiles, logs, integrations, and an API and integration guide so multiple agents can create, update, assign, and complete tasks.

The presenter describes an eight-phase autonomous build process in which agents receive context, plan, create or claim tasks, build, validate, repair problems, report progress, and close the work. The dashboard is intended to support human and agent assignees, subtasks, priorities, due dates, and a “needs input” stage for blocked work.

In the presenter’s tests, OpenClaw’s own cron and session mechanisms did not reliably handle long-running task polling. The workflow became more dependable when a separate local process polled the task API and woke OpenClaw only when work was available. The video recommends using simple scripts or schedulers for dependable polling, while using OpenClaw for reasoning and execution.

AI employees and business workflows

The main employee-style systems demonstrated or discussed include:

• Jane, a meeting assistant: Connects meeting data through APIs and webhooks, stores transcripts and action items, and makes meetings searchable. The presenter says his system had access to more than 1,600 meeting transcripts. One proposed workflow detects discovery calls and creates tasks for a proposal and a prototype.
• Lexa, a voice and phone agent: Uses a voice platform and phone integrations to handle inbound and outbound calls, including speed-to-lead calls after form submissions. The dashboard tracks call logs, transcripts, leads, campaigns, and analytics. The presenter also demonstrates a call function that can send an email during a conversation.
• Nova, an email and SDR system: Supports templates, sequences, campaigns, delivery tracking, and engagement analytics. A prospecting feature turns a natural-language target description into searches, gathers leads through a scraping service, and prepares personalized outreach. The video emphasizes approval and quality controls rather than treating mass sending as the goal.
• Additional business workflows: The presenter discusses competitor research, content operations, lead enrichment, daily business digests, and other specialized agents as extensions of the same architecture.

The examples show how a meeting transcript or submitted lead can trigger downstream work—such as research, proposal preparation, prototype generation, or a phone call—without the user manually initiating every step.

Multi-agent communication and agent-to-agent commerce

The video demonstrates agents communicating through Discord, with individual bots and channels, as well as an agent “council” where agents can discuss a decision and escalate to a human when needed. It also shows OpenClaw and a coding agent handing work to each other, with the coding agent building an application while OpenClaw coordinates and reviews the result.

The final major project is Claw Alley, a marketplace where AI agents list and purchase services. The tutorial uses a web app, Supabase, MetaMask, the Base network, USDC, and an X402-style payment flow. The presenter demonstrates a test and a claimed live transaction between two agents. Buyers need a funded wallet and gas, and the presenter warns against sharing private keys in chat.

Reviews, examples, and limitations

The presenter cites outside examples to argue that these systems can generate business value, including reports of AI-generated proposals, revenue attributed to an AI workflow, and an agent-run content and outreach operation. These are examples cited by the presenter, not independently verified results in the video.

The build process also exposes substantial limitations:

• Configuration changes can break OpenClaw.
• Agents may claim tasks without completing them.
• Front-end and database state can get out of sync.
• Agents may repeat actions or fail to notify the user.

The presenter repeatedly tests, debugs, and revises the integrations. His main practical recommendation is to use clear role separation, persistent memory and logs, explicit task states, secure credentials, and end-to-end verification.

Guides and tutorials covered

The video is structured as a broad, chapter-based tutorial, with sections that can be watched independently. Topics include:

• Installation and local-machine versus VPS setup.
• Sub-agent configuration and token optimization.
• Security hardening and safe API-key handling.
• Business context, memory architecture, and agent identity.
• Mission-control dashboard and Kanban task management.
• GitHub, browser control, email, webhooks, and Supabase integrations.
• Meeting intelligence, voice AI, email campaigns, and lead generation.
• Multi-agent orchestration, Discord communication, and scheduled automations.
• An agent marketplace with on-chain payments.

Main speakers and sources

• Main speaker: Mani Kanasani, the video’s presenter and builder.
• Other sources discussed or demonstrated: OpenClaw and Nvidia NemoClaw documentation or releases; Supabase, Fathom, Resend, voice and phone platforms, and other integration services; and examples attributed to Reddit and social-media posts.

No other clearly identified on-camera speaker appears in the subtitles.