Skip to Content

AI agent for small business: my OpenClaw experience

Several agents with separate permissions, memory that stays in-house, approvals and a tracked budget: what I learned.
November 15, 2026 by
AI agent for small business: my OpenClaw experience
Federico Solutions, Myrko Federico

An AI agent in my small business: an OpenClaw experience report

Field: AI agents / Self-hosting / Security

Period: 2025 to July 2026

Client: Internal project — my own company's AI assistant

Context

For several months, I ran an AI assistant I called "Jarvis" on my own server. It was built on OpenClaw, an open-source agent gateway (MIT licence) that connects a language model to chat channels, tools and memory. I wrote to it from a self-hosted team messaging app. It kept my tasks in my notes, looked things up in my ERP, searched the web and drafted e-mails.

I wanted to answer, on my own company, the question many business owners ask: is an AI agent for a small business a real gain or one more risk? Here is what worked, what got stuck, and what still holds today, now that I have changed tools.

Problem

  • An agent that does everything can reach everything. If it reads e-mail, browses the web and runs commands, a booby-trapped e-mail or a malicious page can feed it instructions. This is called prompt injection: it carries them out with all its rights.
  • It forgets. Without memory, every conversation starts from scratch, and you explain the context again several times a day.
  • It acts without asking. A command on the server or a sent message cannot always be undone.
  • Every wake-up costs money. On each call, an agent sends the model its instructions and the task history. Without control, the bill follows.

Solution

I built the setup in layers, each answering one of these problems. I documented them in the Sovereign AI Protocol guides.

1. Separate agents by trust level

Instead of one all-powerful agent, four agents with separate permissions: a main agent that takes my requests and orchestrates, an operations agent for the server, a communication agent for e-mail and calendar, and a web agent for research. The last two read content from outside, which cannot be trusted: they ran in a disposable container with a reduced set of tools. The web agent could not send e-mail, and the communication agent could not administer the server. For a search, the main agent delegated to the web agent and only received its summary: a hidden instruction on a page stayed inside the container.

2. Give it a memory that stays with me

I added persistent memory hosted on my server: Cognee, an open-source memory engine, with search by meaning and a knowledge graph. Before answering, the assistant retrieved the relevant memories: decisions made, project status, preferences. I no longer re-explained the context every morning, and that history never went to any vendor.

3. Require approval before execution

OpenClaw provides "exec approvals": a per-agent policy that says whether a command is refused, allowed because it is on a list, or subject to human approval. The agents exposed to outside content were not allowed to run commands on the server. The main agent had broad rights: it was my operator zone, a deliberate choice.

4. Keep costs under control

A model chosen for each task, a fast and cheap model for the secondary agents, caching of repeated instructions, pruning of idle sessions and spending monitoring. On my setup, as published on my AI page: about 57 dollars a month in total (7 for the server, 46 for model APIs, 4 for extras), for 200 to 400 requests a day, about 1.5 million tokens a day and three users.

5. Simplify when complexity stops paying off

In May 2026, I reviewed the architecture. The main agent and the operations agent had the same rights and the same rules, so I merged them. I kept the two isolated agents, the only real technical barrier against booby-trapped content. And I planned to remove, agent by agent, access to the ERP and notes connectors: the web agent has no need for them.

Under the hood: for your IT person

OpenClaw is an open-source gateway: a "gateway" process on the host, and agents that each have their own workspace, sessions and tool policy, mapped to channels by routing rules ("bindings"). What running it taught me:

  • Tool policy comes before the sandbox. A tool denied to an agent does not come back through the container. MCP server tools go through the same policy as the others: you can restrict access to them agent by agent.
  • The shell isolation mirage. Denying the browsing tool to an agent that keeps a full shell does not stop it from running curl. The rule "delegate browsing to the web agent" is behavioural guidance, not a barrier. Only the container and the absence of a shell make it one.
  • The Docker sandbox is not a fortress. The documentation says so itself: it is not a perfect security boundary, and the gateway stays on the host. Read-only or no workspace, exec denied. A custom image (mail client, calendar) has to be rebuilt with every update.
  • Exec approvals. Modes deny, allowlist or full, ask off, on-miss or always, and refusal by default when no interface answers. Approvals can only tighten the configuration, never loosen it.
  • Memory. Natively, Markdown files in the workspace and hybrid search. I added Cognee (pgvector, graph) with automatic recall before each answer. Before compacting the history, OpenClaw asks the agent to write what matters to disk.
  • The real failures. A new session took 30 to 60 seconds to start. A compaction handed to a remote model that timed out blocked the gateway and dropped the channel. The prompt cache collapsed whenever the system prompt changed.
  • Checks. The built-in security audit command flags configuration drift: run it after every change.

Result

  • For several months, an assistant reachable from my phone that remembered my projects and worked inside my tools.
  • A measured and published cost: about 57 dollars a month for three users.
  • A method documented in ten guides, including a free introductory guide.
  • A change of course: in July 2026, I stopped my OpenClaw gateway to run other trials, around Claude Code, Anthropic's official tool, as I explain in claudeVPS. OpenClaw reused my Claude subscription in a way that Anthropic's terms regulate ever more strictly. The foundation has not changed: the secure server, memory and monitoring are still running and carry these new trials. That foundation is what my guides document, and it holds whatever agent you install on it.

What I take away

  • Separating beats watching. An agent that reads the web must not be able to write to the ERP. I keep that separation whatever the tool.
  • An instruction is not a barrier. What protects you is what the agent technically cannot do.
  • Memory makes the assistant. It turns a chatbot into a colleague, and it should stay with you.
  • An agent is software to maintain. Updates, migrations, outages: budget for operating time, not just tokens.
  • The principles outlive the tool. Separate permissions, memory, approvals, budget: I carried them over unchanged to my current setup. I detail the guardrails in this article, and how to budget your agents in this one.

Going further

The architecture is described step by step in my AI server guides: guide 6, multi-agents and approvals, guide 5, persistent memory, guide 7, cost control. Guide 0 is free: architecture, costs and required skill level. Before connecting an agent to your ERP, also read MCP server security. To discuss your case: a 20-minute discovery call.

The step-by-step guides

Each layer described here has its own guide, at CHF 39: no. 4 for the team chat connected to the assistant, no. 5 for persistent memory, no. 6 for agents with separate permissions and approvals. The Guides Pack (CHF 149) brings together guides 0 to 8: a complete stack, documented and tested in production, from the secure server to backups, secrets and monitoring.

Sources

  • openclaw/openclaw, OpenClaw, GitHub, MIT licence, accessed 15 November 2026
  • Multi-Agent Routing, OpenClaw, documentation, accessed 15 November 2026
  • Sandboxing, OpenClaw, documentation, accessed 15 November 2026
  • Exec approvals, OpenClaw, documentation, accessed 15 November 2026
  • Memory, OpenClaw, documentation, accessed 15 November 2026
  • MCP, OpenClaw, documentation, accessed 15 November 2026
  • Security, OpenClaw, documentation, accessed 15 November 2026

Cover photo: Pixel.la Free Stock Photos / Wikimedia Commons, public domain (CC0).

Tags: AI agent, small business, OpenClaw, multi-agent, persistent memory, prompt injection, LLM cost, self-hosting

3-2-1 backup: copies that actually restore
The 3-2-1 rule, encryption, retention and restore testing: what separates a reassuring copy from one that saves you.