Two open-source projects have shaped how people think about AI agents in 2026. OpenClaw showed how much a self-hosted agent can do when you connect a capable model to everyday tools and talk to it through a chat app. Hermes Agent, from Nous Research, pushed further on memory and on agents that improve by turning experience into reusable skills. Both are free, both are popular with developers, and both raise the same question for businesses: how do we use this safely? This guide explains what each is and how we deploy them for clients.
If you are new to agents, start with What is agentic AI?
What OpenClaw is
OpenClaw is a self-hosted AI agent runtime. You install it on a server, a workstation or a small device, connect it to a model provider (commercial or local), and give it instructions through messaging apps such as Slack, Telegram or WhatsApp. Rather than only replying, it acts: it can read and write files, run commands, browse the web, send email and manage calendars, and it can run tasks on a schedule. It started in late 2025 under a different name and is now maintained by an independent open-source foundation.
What Hermes Agent is
Hermes Agent is an open-source agent framework from Nous Research, released under the MIT licence. It runs locally or in the cloud, works with many model providers including local models, and connects to a wide range of messaging platforms from one gateway. Its distinguishing idea is a learning loop: persistent memory across sessions and the ability to create and refine reusable skills from what it has done before.
How they differ
| Area | OpenClaw | Hermes Agent |
|---|---|---|
| Origin | Independent open-source project, now foundation-run | Nous Research |
| Core idea | A capable personal agent on your own hardware | An agent that learns skills and keeps long-term memory |
| Interface | Chat apps | Chat apps and other integrations |
| Models | Commercial or local | Commercial or local |
| Main risk | Broad system access | Broad system access |
What businesses use them for
- Internal assistants that research, summarise and draft across shared drives and knowledge bases
- Operations helpers that monitor inboxes, triage requests and prepare responses for approval
- Scheduled reporting that gathers data from several tools and posts a summary to a team channel
- Developer and IT automation such as log review, ticket enrichment and routine maintenance tasks
- Fast prototypes that show which workflows are worth turning into a purpose-built agent
An open-source agent installed with default permissions is a new employee with every password and no manager. The value is real, but only once the guardrails are in place.
ESS ENN Associates AI Engineering Team
The security problem, stated plainly
The feature that makes these agents useful — they can act on real systems — is also the risk. An agent with access to files, email and a shell, that also reads untrusted content from the web or incoming messages, can be manipulated through prompt injection into leaking data or taking actions nobody intended. Credentials stored for convenience become an attack surface. Security researchers have flagged these issues for both projects, and both recommend running them carefully.
How to deploy them safely
- Isolate it. Run the agent in a sandboxed container or dedicated virtual machine, never on a laptop holding company credentials.
- Give it its own identity. Separate service accounts with least-privilege access to only the systems each workflow needs.
- Gate sensitive actions. Require human approval for sending external messages, spending money, deleting data or changing permissions.
- Filter what it reads. Treat web pages, emails and documents as untrusted input, and restrict which tools can act on them.
- Log everything. Keep a full trace of prompts, tool calls and results, and alert on unusual behaviour.
- Control the model. Choose commercial or local models deliberately; for sensitive data, use self-hosted open-weight models so nothing leaves your environment.
- Review skills and plugins. Vet third-party skills before installing them, as you would any other software.
Open-source agent or custom agent?
Open-source agents are an excellent way to automate personal and team workflows quickly. For customer-facing processes, regulated data or business-critical operations, a purpose-built agent with tightly scoped tools, evaluation and monitoring is usually the better fit. We cover that path in deploying custom AI agents to production and custom AI agents.
How we help
ESS ENN Associates deploys, customises and secures open-source agents such as Hermes Agent and OpenClaw: hardened self-hosted installations, custom skills and tool integrations for your systems, approval workflows, logging and ongoing operation. For organisations that need everything to stay in-house, we run them against private models on on-premise or air-gapped infrastructure. Our agent practice operates as No Ifs, Only Bots, and builds custom agents through our AI agent development service. To discuss a deployment, contact us.