Self-Hosted OpenClaw and Hermes Agent for Companies
Self-hosted AI agent deployment for companies that want an always-on assistant for the team, on servers they control. We run OpenClaw or Hermes Agent as private AI agents in an isolated VM or container in your cloud account, with scoped service accounts, approved skills and full logging. Pair them with a hosted model, or run them as local AI agents on a model inside your own network.
Why Companies Run a Self-Hosted AI Agent
A personal AI agent such as OpenClaw or Hermes Agent reads email, opens files and joins team chat, so it holds the same data your staff do. Its state, memory and credentials live on the machine it runs on, so companies run it on machines they own:
A private AI agent your team shares, inside your perimeter.
We deploy OpenClaw or Hermes Agent as a self-hosted AI assistant for the team, in the chat tools you already use, with every account scoped and prompts sent only to the model you choose.
OpenClaw, Hermes Agent and NVIDIA NemoClaw
What each project is and its licence, from its own repository and docs. Either agent can also run inside NemoClaw.
OpenClaw
An open-source AI assistant that answers in Slack, Teams, WhatsApp, Telegram, Discord and 20+ more channels. One Gateway serves one person or a whole team. MIT licence, stewarded by the independent OpenClaw Foundation.
Hermes Agent
Nous Research's self-improving agent. It writes skills from completed tasks, keeps memory across sessions, runs scheduled jobs and talks through Telegram, Discord, Slack, WhatsApp and Signal. MIT licence.
NVIDIA NemoClaw
NVIDIA's open-source reference stack, announced in March 2026, runs OpenClaw or Hermes in an OpenShell sandbox with network policy and credentials kept outside. Apache-2.0 licence; NVIDIA calls it an alpha project.
An Isolated VM or Container in Your Cloud Account
The agent gets its own machine, network rules and identities, inside the cloud account or data center you already run.
We deploy the OpenClaw or Hermes Agent gateway on a dedicated VM or container in your AWS, Azure or Google Cloud account, or on your own servers, segmented from production systems. Shell commands run in a Docker, OpenShell or SSH sandbox, depending on the agent.
OpenClaw binds to loopback on a regular host install, per its security docs, and staff reach it through Tailscale or an identity-aware proxy. Teams with different data rights each get their own gateway. The agent can sit inside a wider self-hosted AI deployment with private chat and document search.
OpenClaw Security Hardening for Company Deployments
OpenClaw security drew wide attention in early 2026. SecurityScorecard's STRIKE team counted more than 40,000 internet-exposed instances in its first 24 hours of scanning, CVE-2026-25253 (CVSS 8.8, fixed in 2026.1.29) let a crafted link leak an instance's gateway token, and Bitdefender Labs flagged about 17% of the OpenClaw skills it reviewed as malicious. Our deployments add these controls, which cover Hermes Agent security too.
A private Gateway with authenticated access
The Gateway listens on loopback or a private subnet, and staff reach it through SSO-backed ingress.
Current releases and audits
We pin a current release, keep an update routine and run openclaw security audit after changes. Hermes Agent uses hermes update and its production checklist.
Sandboxed tools
OpenClaw runs tools on the host until sandboxing is on, so we enable Docker or OpenShell. Hermes Agent runs commands in its Docker backend, with a non-root gateway.
Allowlisted people and rooms
Only named users and channels trigger the agent, and unknown direct messages get a pairing code.
Approval for risky commands
Destructive commands wait for a person, through Hermes Agent's approval modes and OpenClaw's exec approvals.
Limits on prompt injection
Email, web pages and shared files count as untrusted input, so agents that read them get the narrowest tool set.
Per-Agent Accounts, Read-Only Defaults and Approvals
The agent acts with the accounts you give it, so each one is scoped to the job.
Per-agent service accounts
Each agent gets its own mailbox, drive and chat identities, reviewed like any service account. On NemoClaw, OpenShell adds credentials only to requests for approved endpoints.
Read-only by default
Email, files, CRM and calendars start read-only. Write access is added per task, for the folders and records that task needs.
Approvals for sends and payments
The agent drafts outgoing email, customer messages and payments, and a named person approves each one in chat before it goes out.
An Approved Skills List and MCP Access Through a Gateway
Skills and tools decide what the agent can do, so the company decides which ones load.
Approved skills per agent
OpenClaw loads only the skills on each agent's allowlist. We set that list per agent and keep it in version control. Hermes Agent writes skills from completed tasks, and we review those before they join the shared set.
Review of third-party skills
We read every ClawHub or public skill before install, scripts and network calls included. ClawHub shows VirusTotal, ClawScan and static analysis results, and openclaw skills verify checks a skill against the registry.
MCP access through the gateway
Company systems reach the agent as MCP tools behind a self-hosted MCP gateway, with SSO, per-tool scopes and an audit log. The approved agents and skills go into your AI acceptable use policy.
A Hosted Frontier Model or a Local Model on Your GPUs
Both agents support hosted model APIs and self-hosted models, per their provider docs. We match the model to the data each agent handles.
Hosted frontier model
Anthropic, OpenAI or Google models under your own API keys, routed through a self-hosted LLM gateway that holds the keys, sets per-agent budgets and logs each call.
Local model on vLLM
Open-weight models served by vLLM on your GPU servers, through the OpenAI-compatible API that both agents can call.
Local model on Ollama
A local AI agent for smaller teams. OpenClaw connects through Ollama's native API, which keeps tool calling intact, and Hermes Agent supports Ollama too.
Agent Logs Your Security Team Can Query
Every message, tool call and model call leaves a record in the systems your security team already uses.
Agent logs and OpenTelemetry
OpenClaw writes JSON-lines logs and exports OpenTelemetry metrics, traces and logs. Hermes Agent keeps its own log files. Both ship to central storage.
SIEM forwarding and alerts
Commands, API calls and approvals go to your SIEM, as TechTarget's guidance for CISOs recommends, with alerts on new skills.
Model calls in the LLM gateway
Each prompt and response is logged with the agent, model and cost, so spend per agent stays visible.
One-Click OpenClaw Hosting Compared With a Company Deployment
Hostinger, DigitalOcean, Contabo and LumaDock offer one-click OpenClaw installs, per Hostinger's 2026 round-up of OpenClaw VPS hosts, and Hermes Agent runs on a VPS through its own installer. A company deployment adds what a shared team agent needs.
| Area | One-click VPS instance | Company deployment |
|---|---|---|
| Where it runs | A VPS in the host's cloud, on one account | A VM or container in your own cloud account |
| Sign-in | The owner, with a gateway token or DM pairing | Staff through SSO-backed ingress |
| Credentials | Keys and logins the owner adds | Per-agent service accounts, read-only by default |
| Skills | Whatever the owner installs | An approved, reviewed list per agent |
| Models | Hosted APIs; Hostinger's plan includes AI credits | Your LLM gateway, or a local model on vLLM or Ollama |
| Updates | The owner, on a self-managed VPS | A set routine, with an audit after each change |
| Logs | On the instance | Central logs, OpenTelemetry and SIEM alerts |
What Shapes a Self-Hosted AI Agent Engagement
The scope of an OpenClaw enterprise or Hermes Agent deployment follows the agent's job, the systems it touches and your data rules. Bring the tasks you have in mind, and we will map the deployment that fits.
Book a Strategy Session →Frequently Asked Questions on Self-Hosted AI Agents
It can be, with the right setup. OpenClaw's docs say a regular host install binds the Gateway to loopback and answers unknown senders with a pairing code, and the project publishes fixes on GitHub. The incidents reported in 2026 involved exposed instances, unpatched versions and malicious skills, which the controls above address.
Both are MIT-licensed and self-hosted. OpenClaw connects to Slack, Teams, WhatsApp and 20+ more channels, and has a team mode with shared sessions and roles. Hermes Agent is built around a learning loop that writes and refines its own skills. We help you pick by the chat tools your team uses.
A one-click OpenClaw VPS or Hermes Agent VPS gives one person a running agent in a hosting provider's cloud. A company deployment runs in your own cloud account, with SSO-backed access, scoped service accounts, approved skills, central logs and the model you choose.
OpenClaw and Hermes Agent are free under the MIT licence, and NemoClaw under Apache-2.0. Running costs are the VM, model usage or GPUs for a local model, and log storage. A strategy session sizes these for your team.
Related Self-Hosted AI Solutions
Self-Hosted MCP Gateway for Claude and ChatGPT Connectors
Learn more →Self-Hosted LLM Gateway for Model Routing and Cost Control
Learn more →Self-Hosted AI Deployment: A Private AI Stack in Your Environment
Learn more →Ready to Run a Self-Hosted AI Agent for Your Team?
A free strategy session on the tasks you want the agent to handle, the systems it would touch and your data rules, with a sketch of the deployment that fits.