Hermes Agent Backend: Every Type Explained, And The Models Worth Putting Behind It

A Hermes Agent backend is the layer that does the work behind the chat, and the main one is the terminal backend, which decides where the agent's shell commands run.

Hermes Agent is the free, open-source AI agent from Nous Research, and its docs list seven terminal backends: local, Docker, SSH, Modal, Daytona, Vercel Sandbox and Singularity.

The word "backend" is also used for the model provider, the web search provider, the browser provider, the memory provider, and Hermes itself running as a backend for other apps.

I took every backend fact on this page from the official Hermes docs and the read-only help output on my own install, which is version 0.21.4.

Below you'll get each backend type, the command that sets it, the security and cost trade-offs, and then a later section on which model to run behind it.

Hermes Agent backend types at a glance

Hermes is released under the MIT licence, so the software is free and the only costs come from the services you plug into those slots.

The terminal backend: where commands run

The docs say each terminal backend determines where the agent's shell commands actually execute.

The default is local, and my own config is still set to it.

How to switch your Hermes Agent backend

You check the current terminal backend with hermes config get terminal.backend.

You change it with hermes config set terminal.backend docker, and you swap in any of the other six names.

The value is saved in ~/.hermes/config.yaml under the terminal: block, so you can edit the file by hand too.

If a new backend fails, you set the value back to local and run hermes doctor.

How each terminal backend behaves

Local

Commands run with your real home folder, so tools like git, gh and npm find the logins they already use.

The docs warn that the agent has the same filesystem access as your user account.

Docker

Hermes starts one long-lived container on first use and sends every terminal, file and code call into it with docker exec.

A package the agent installs once is still there on the next command, and it's still there after you restart Hermes.

Setting container_persistent: false gives every chat its own fresh container.

Setting docker_network: false runs the container with no network at all.

Container backends default to 1 CPU core and 5 GB of memory, and you can raise both.

SSH

Hermes keeps one shell alive on the remote host, so the working directory and environment variables survive between commands.

The tools page recommends SSH for security, because the agent can't modify its own code from a different machine.

Modal, Daytona and Vercel Sandbox

These three are cloud sandboxes, and each needs its own account.

They save your files between sessions when persistence is on, and the docs say they don't save running processes.

Daytona caps disk at 10 GiB, and Vercel Sandbox supports the node24, node22 and python3.13 runtimes.

Singularity

This backend exists for HPC clusters and shared machines where Docker isn't available.

The model backend: which provider does the thinking

The second meaning of a Hermes Agent backend is the model provider.

You add a provider with hermes model in your terminal, which is the full wizard for logins and keys.

You switch between providers you've already added with /model inside a chat.

The docs call Nous Portal the recommended way to run Hermes, because one login covers 300+ models and the Tool Gateway.

The docs warn that Ollama starts with a small context window, so you set context_length for a local model.

You can add a backup chain with hermes fallback, which switches provider when the main model fails.

The AWS route has its own page, Hermes Agent on Bedrock.

Web, browser and memory backends

Hermes has two web tools, web_search and web_extract, and Firecrawl is the default backend for both.

The docs also list SearXNG, Brave Search, DuckDuckGo, Exa, Parallel, Tavily, Perplexity, Keenable, xAI and OpenAI Native.

A fresh install works with no key, because requests rotate across the free tiers of Exa, Parallel, Firecrawl and Keenable.

You can split the work with web.search_backend and web.extract_backend.

The browser toolset has its own backends, which are Browser Use, Browserbase, Firecrawl, Camofox, Lightpanda, your own Chrome over CDP and a local Chromium.

I explain that toolset in the Hermes Agent browser guide.

Memory has a built-in layer in MEMORY.md and USER.md, and that layer is always on.

You can add one external provider with hermes memory setup, and I rank them in the best Hermes Agent memory.

Using Hermes Agent as a backend

The last meaning flips the direction, because Hermes becomes the backend for another app.

The API server speaks the OpenAI Chat Completions and Responses formats, and the docs name Open WebUI, LobeChat and LibreChat as frontends.

The agent answers with its full toolset, so the frontend gets terminal, files, web search, memory and skills behind it.

The hermes serve command binds to 127.0.0.1 by default, and a public bind always requires an auth provider.

Security trade-offs: local versus Docker

On local, Hermes checks commands against a list of dangerous patterns and asks you before it runs a match.

On Docker, the container starts with all Linux capabilities dropped, no privilege escalation and a limit of 256 processes.

The docs recommend docker, modal, daytona or vercel_sandbox for production gateway deployments.

A prompt only protects you when a human is there to read it, and a container protects you when nobody is watching.

Any secret you list in docker_forward_env can be read by code inside the container, so forward as few as you can.

Performance and cost

The local backend has no sandbox to start, so nothing sits between the agent and your shell.

The Docker backend reuses one running container, and the SSH backend reuses one connection and one shell.

Every command has a 180 second timeout by default.

The docs warn that slow shell startup files add delay to every command, because the agent starts many shells.

Local, Docker, SSH and Singularity cost nothing beyond your own hardware or server.

Modal, Daytona and Vercel Sandbox are paid cloud services with their own pricing, and your model provider bills separately.

If you want the agent to keep working on these backends after you log off, read Hermes Agent background tasks.

Which model to put behind your Hermes Agent backend

Hermes Agent doesn't ship its own model, so the model backend you pick decides how good the agent is.

I use GoldieBench to shortlist that model, and I want to be straight about what the board measures.

GoldieBench gives each model the same one-shot prompt to build a game, a page, a simulation or a visual, then scores the rendered result from 0 to 10.

It doesn't test terminal backends, Docker sandboxes or tool calling, so there is no GoldieBench score for local versus Docker.

The scores do show how well a model handles hard, multi-step build work, which is the closest public signal I have for an agent's brain.

Those numbers are the live averages on each model page when I checked on 12 October 2026.

How I'd pair a model with each backend

A heavy terminal job inside Docker involves many decisions in a row, so I'd put a frontier model such as Claude Opus 5 behind it.

A cheap job that runs all day is a better fit for a lower-cost model such as MiniMax M3.

A private job on the local backend pairs naturally with a local model from the local board, such as Qwable 5 27B Coder.

The live page describes Qwable as MLX-only, so it reaches Hermes through a custom OpenAI-compatible endpoint and not through a hosted provider.

A local model is free to run, and it's slower and weaker than the frontier models in that table.

My wider picks are in the best LLMs for Hermes Agent and the best local model for Hermes Agent.

My simple decision rule

I'd stay on local while I'm the only person who talks to the agent.

I'd move to Docker the day a second person can message it.

I'd use SSH when I want the agent on a server that never sleeps.

I'd pick the model after the terminal backend, because safety comes before speed.

The rest of my stack is in the best Hermes Agent setup, and the toolsets that sit on top are in the best Hermes Agent tools.

Also On Our Network

• Agent Operatorsthe exact commands for switching every Hermes backend
• AI Income Deskwhich Hermes backend a business should run
• AI Tool Verdictall seven Hermes terminal backends scored
• agentos.guidea quick-start for Hermes backends inside an Agent OS
• aiprofitboardroom.comJulian's Hermes Agent Docker walkthrough

Pick the room first and the brain second, and you'll have the right Hermes Agent backend.