n8n, Make, or Claude: What For?
n8n vs. Make: prices, hosting, and GDPR compared, plus the question rarely asked: do you need a workflow builder if Claude can access mail directly?
Read the articleNovember 1, 2026 · 14 min read
The right OpenClaw alternative for a business depends on the job: Claude Code for an agent that works in files and code with permission rules, Claude Cowork for desktop work without a terminal, n8n or Make for fixed workflows, and Hermes Agent if you want another open-source agent you reach through a messenger. OpenClaw itself is an open-source AI assistant under the MIT license that runs on your own machine and answers in WhatsApp, Telegram, Slack, Teams and, according to the project, more than 20 other channels. A local service, the Gateway, connects them to a model (Claude, Codex or a local one), tools, skills and plugins. The risk: according to the docs, tools run directly on the host unless you configure a sandbox, plugins run fully trusted, and any message or web page the agent reads can carry a prompt injection. If you keep OpenClaw, run it on a dedicated machine or VM without personal accounts, keep the Gateway on loopback, turn on the sandbox and read skills and plugins before installing them.
I work with Claude every day and run my team of fifteen AI employees on it. OpenClaw was not installed or tested for this article. Everything here is based on the project's repository, docs and security advisories, retrieved on November 1, 2026; the current version was 2026.10.9 from October 29, 2026. Everything I have written about AI agents is collected on my AI agents page.
Peter Steinberger built OpenClaw with the community; today it is stewarded by the OpenClaw Foundation, a US nonprofit. According to the README there is no paid tier and no hosted service; OpenAI is among the donors, in the project's words "a donor, not an owner". On November 1, 2026 the GitHub repository had about 394,000 stars.
AI agents in general are explained in What Is an AI Agent? Explained Simply. What sets OpenClaw apart: you message it like a colleague, and it acts on your computer with your accounts.
SKILL.md that teach the agent when to use which tool. Plugins are code packages for new capabilities, channels or model providers.At the centre sits the Gateway, a long-running service that holds the channels, manages sessions and tools and listens on port 18789 by default, on a regular host only on 127.0.0.1 according to the docs.
| Component | What it is | Where the risk lies |
|---|---|---|
| Gateway | local service, port 18789 | whoever has access controls the agent completely |
| Channels | messenger connections | anyone allowed to write shares the agent's permissions |
| Tools | shell, files, browser, web | run on the host without a sandbox |
| Skills | Markdown instructions | third-party skills are "untrusted code" per the docs |
| Plugins | code in the Gateway process | run fully trusted |
~/.openclaw |
settings and state | tokens, messenger credentials, transcripts |
Skills and plugins come from ClawHub, the project's public registry. Per the docs anyone with a sufficiently old GitHub account can publish there; ClawHub scans new releases and shows the result before you install. According to the README, OpenClaw by default only phones home for a daily version check.
The threat model in the docs opens with: "Your AI assistant can execute arbitrary shell commands, read/write files, access network services, and send messages to anyone (if given channel access)." Five risks follow from that.
1. Permissions. Without a sandbox, tools run directly on the host, for a single operator by default without approval prompts, by design. The agent can do whatever your user account can do.
2. Prompt injection. Any content the agent reads, whether a web page, an email or an attachment, can carry instructions: "The content itself is a threat surface". The project's security policy does not count prompt injection on its own as a vulnerability. Limiting the damage is your job. How these attacks work is covered in Prompt Injection: Protecting AI Agents.
3. Plugins and skills from third parties. "Plugins run in-process with the Gateway - treat them as trusted code", the docs say. Installing one runs code, and per the docs OpenClaw does not block dangerous code on its own; you set up an install policy yourself. Under its security policy, a malicious plugin you installed is not an OpenClaw vulnerability.
4. Open ports. In a container the Gateway listens on all addresses (0.0.0.0) by default, according to the docs. It will not start there without authentication, but whoever holds the token has full operator access.
5. Credentials. ~/.openclaw holds tokens, messenger credentials, model provider profiles and transcripts. The docs recommend permissions 700 and 600, full-disk encryption and a dedicated user.
Between January 31 and September 11, 2026 the project published 722 security advisories on GitHub, 14 rated critical, such as a path traversal in plugin installation (fixed in 2026.2.1) and a sandbox escape (fixed in 2026.3.31). None was added up to the retrieval on November 1, 2026. Keep it updated anyway: every older version still carries the holes fixed since.
No ranking, because the tools solve different jobs:
In my own setup, five of my AI employees run with Claude Code on a small Linux server, each role has its own Slack bot, and each bot only accepts jobs and approvals from my own Slack account. The server has no open port, and credentials sit encrypted in the repository. I describe the setup in Claude Code on a Server: tmux, Remote Control. How Claude itself connects to services such as mail or calendar is covered in Claude Connectors: What They Are and Do. More tools are compared in Claude Code Alternatives in 2026.
For a business, OpenClaw is not a vendor product but software you run yourself: the Gateway runs on your side, your input goes to whichever model provider you connect, and updates, permissions and reviews are up to you.
| Situation | Assessment based on the docs |
|---|---|
| One technically skilled person tries it on a separate VM with test data | fits the docs' model: one operator, separate environment |
| Several colleagues message the same agent | per the docs, everyone allowed to message it "shares that agent's delegated tool authority" |
| The agent sees client data, mailboxes or payment accounts | any email or web page it reads can inject instructions; without a sandbox and limited tools, not a good place |
| Nobody is in charge of updates | with 722 published advisories in just over seven months, a real risk |
This isn't legal advice, just the operator's perspective. As soon as personal data passes through the agent, you sort out data protection with the model provider and for your own server, as with any other AI tool. Which data an AI should see at all is covered in AI and Privacy: What the AI Gets to See.
Before real data flows, a trial run is worth it, on a throwaway VM without your credentials and with a separate API key and spending limit. Put a test file with a hidden instruction there and see whether the agent follows it, once with defaults and once hardened with the sandbox.
npm install -g openclaw@latest. Download the script first (curl -fsSL https://openclaw.ai/install.sh -o install.sh), read it, then run it with bash install.sh.gateway.bind: "loopback" explicitly, use token auth, and per the docs prefer Tailscale Serve over a LAN bind for remote access.agents.defaults.sandbox set to mode: "all", every tool, including those of your own session, runs in a Docker or Podman container and cannot see your workspace by default (workspaceAccess: "none"). The docs: "This is not a perfect security boundary". Leave tools.elevated off.plugins.allow), pinned versions, read the code before enabling.openclaw security audit reports drift from the safe defaults. Then update regularly.For agents with tools, the docs recommend the strongest current model, because smaller ones are easier to manipulate.
An open-source AI assistant under the MIT license that runs on your own computer and is reachable through messengers such as WhatsApp, Telegram or Slack. A local Gateway connects channels, model, tools, skills and plugins.
The defaults are cautious per the docs: the Gateway listens locally only, and unknown senders have to pair. But without a sandbox tools run directly on the host, and prompt injection is not ruled out. It gets safer with a separate environment, a sandbox, limited tools and updates.
The software is; according to the README there is no paid tier. Costs come from the model and the machine the Gateway runs on.
Yes. According to OpenClaw's docs either with an Anthropic API key, billed by usage, or through an existing Claude Code login on the same machine, in which case usage counts against your subscription. For shared automation the docs recommend the API key.
OpenClaw's public registry for skills and plugins. New releases are scanned automatically, but the docs still treat third-party skills as untrusted.
It depends on the job, see the list above. More important than the tool is that permissions, sandbox and approvals are enforced technically.
Before installing OpenClaw, decide where it runs: on a dedicated machine or VM, not your work computer. Start with the hardened baseline and the sandbox, and run openclaw security audit before connecting a messenger. If you would rather work with Claude, start with Install Claude Code: Windows, Mac, Linux. How I build AI employees with clear permissions and approvals is something I show in my community AI Employee Framework.
n8n vs. Make: prices, hosting, and GDPR compared, plus the question rarely asked: do you need a workflow builder if Claude can access mail directly?
Read the articlen8n and Make are the standard advice for AI agents. Why I take a different path: managing employees instead of building workflows.
Read the article
Kevin Welter
Developer, IT architect, author of technical books (Kubernetes, cloud infrastructures) and speaker. Runs his business with an AI workforce of fifteen AI employees and shows solo business owners in his community how to hire their first AI employee.
Your first AI employee up and running within an hour
Join the community$5/month founding price, first 50 members