Hermes Agent, the open-source self-improving AI assistant from Nous Research, just became fully portable across any model provider and execution environment. On August 7, Vercel announced that Hermes can now use AI Gateway as its inference layer and run every agent command inside a Vercel Sandbox microVM, completing a pattern where both the model layer and execution layer are entirely configurable by the user.

What Just Changed for Hermes Users?

The integration adds two independent, opt-in switches that developers can configure from the command line. First, Hermes now appears in the AI Gateway setup wizard, allowing it to route through over 200 models on the gateway with no token markup and unified spend tracking across all model vendors. Second, developers can now point Hermes to Vercel Sandbox, which isolates each agent command in a cloud microVM instead of running it locally on their machine.

The practical setup is straightforward. Existing Hermes users can switch to the new backends with three commands: hermes update, hermes setup model to pick Vercel AI Gateway, hermes setup terminal to pick Vercel Sandbox, and hermes doctor to verify the configuration. This makes Vercel Sandbox the seventh terminal backend option for Hermes, joining local execution, Docker, SSH, Singularity, Modal, and Daytona.

How Does Hermes Actually Work as a Self-Improving Agent?

Hermes is built around a self-improving loop that sets it apart from standard chatbots. When the agent solves a complex task, it creates reusable skills from that experience, refines them during use, and maintains persistent memory across sessions. This means the agent gets faster and more tailored to your specific workflows over time, rather than forgetting everything after each conversation.

The agent supports autonomous goal-based work. Instead of having a conversation, you define a mission like "Organize my downloads folder by file type every day at 9 AM." Hermes then breaks the goal into steps, executes them, and learns from any obstacles it encounters. It can browse the web, schedule automations, run subagents, and execute shell commands, making it a general-purpose autonomous tool rather than a specialized chatbot.

Steps to Get Started With Hermes Agent

• Set Up Your Infrastructure: Install Hermes using the official installer (curl -fsSL https://raw.githubusercontent.com/NousResearch/hermes-agent/main/scripts/install.sh | bash) on Linux, macOS, WSL2 for Windows, or Android via Termux. Alternatively, use Docker for a containerized setup that stays consistent across local and cloud environments.
• Configure Your Model Provider: Obtain an API key from an LLM provider like DeepSeek, OpenAI, Claude, or a local Ollama endpoint. OpenRouter offers a free key as the fastest starting point. Then run hermes setup model to route through Vercel AI Gateway and choose from 200+ available models.
• Choose Your Execution Environment: Decide whether to run commands locally on your machine, in Docker, or in a Vercel Sandbox microVM. For security and 24/7 uptime, a cheap $5-10 per month VPS or Vercel Sandbox is recommended over running on your laptop.
• Connect a Messaging Interface: Link Hermes to Telegram, Slack, or WhatsApp so you can message the agent from your phone and give it tasks while it runs in the background.
• Start With One Simple Goal: Begin with a single recurring task, like a daily inbox summary, rather than hooking up every tool and model provider at once. Once that workflow is solid, add complexity.

Why Sandboxing and Provider Neutrality Matter

Running an agent's commands in a cloud microVM instead of your local filesystem is a meaningful security improvement. Hermes is a general agent that can browse, schedule automations, and execute shell commands, so pointing it at your local filesystem carries real risk. A Vercel Sandbox microVM that exists only during command execution and can be revoked via OIDC token provides the containment layer most agent setups are missing. The cost is opt-in and minimal, since microVMs only spin up while the agent is working with no idle baseline.

Provider-agnostic routing through AI Gateway removes the last reason to hardcode a single model vendor. Developers get one dashboard for model choice, latency, and cost across all agent traffic, and can switch models with a configuration change rather than rewriting code. When combined with Vercel's spend-budget scoping feature, developers can set hard dollar limits that reject requests if they exceed a threshold, turning cost guardrails into account-level controls rather than per-tool settings.

What Does This Signal About the AI Agent Market?

The consolidation of agent infrastructure is accelerating. Cloudflare unified Workers AI and AI Gateway into a single control plane on the same day Vercel shipped this integration, signaling that the winning position in the agent stack is not the agent itself, but the control plane it runs on. For developers, this means switching between agents becomes cheaper than switching between gateways, and switching models is now a simple configuration change in both major cloud platforms.

Hermes is also part of the broader skills ecosystem, where agents auto-generate skills from solved problems and support the agentskills.io open standard. This design pattern means portable skills define what the agent can do, a neutral gateway defines what it can call, and a sandbox defines where it runs. None of the three layers depends on a specific vendor, creating a modular, composable agent stack.

Common Questions About Running Hermes

Users often ask whether they need to be developers to use Hermes. The answer is no, but you do need to be comfortable with command-line interfaces and basic server configuration. Hermes is open-source under an MIT license, so your prompts, memory, and task data stay entirely on your machine; nothing is sent to Nous Research. Updates are simple: just run hermes update and the tool handles backup and upgrade automatically.

The software itself is free, but costs come from hosting (if you use a VPS or Vercel Sandbox) and LLM tokens consumed by your model provider. You can run Hermes very cheaply using budget models, and it pairs perfectly with Ollama for a completely private, offline-capable setup. If the agent gets stuck, hermes doctor checks the status, and you can interrupt it with Ctrl+C or run /compress to clear long-term context. Hermes also supports fallback providers, so if your main model provider goes down, it can automatically rotate to a secondary one.

Even resource-constrained hardware works. Hermes runs on a Raspberry Pi with an SSD for better performance, making it a great low-power option for 24/7 operation. For those looking to monetize, automating recurring business tasks like lead generation, content distribution, or data monitoring can be sold as a service or used to scale your own business.