Hermes Agent backend is a term with six meanings, 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.

My verdict is that local is the right place to start, Docker is the right place to end up for most people, and SSH is the pick when you want the agent on a server.

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

This page is a reference, so each backend type gets its own section with the options, the setting and my score.

I scored the options from what the official docs say about isolation, setup and cost, and I didn't run a lab test on each one, so treat the numbers as an informed opinion.

Hermes Agent Backend Types At A Glance

Backend type What it decides Documented options How to set it
Terminal It decides where shell commands run. The options are local, docker, ssh, modal, daytona, vercel_sandbox and singularity. You run hermes config set terminal.backend <name>.
Model It decides which model thinks. The options include Nous Portal, OpenRouter, Anthropic, OpenAI, Bedrock, Ollama, LM Studio and custom endpoints. You run hermes model.
Web It decides who serves search and extraction. The options include Firecrawl, SearXNG, Brave, DuckDuckGo, Exa, Parallel, Tavily, Perplexity and Keenable. You run hermes tools or set web.backend.
Browser It decides which browser the agent drives. The options include Browser Use, Browserbase, Firecrawl, Camofox, Lightpanda and local Chromium. You run hermes tools and open Browser Automation.
Memory It decides where extra long-term memory lives. The options include Honcho, OpenViking, Mem0, Holographic, RetainDB, ByteRover and Supermemory. You run hermes memory setup.
Hermes as a backend It decides how other apps reach your agent. The options are the API server, ACP, MCP serve and hermes serve. You run the matching command.

The Seven Terminal Backends, Scored

These scores are out of 10, and they're my opinion from reading the docs on isolation, setup effort and cost.

Backend Safety Ease of setup Cost My verdict
local 4 10 10 It's the best starting point for one trusted user.
docker 9 7 10 It's the best all-round sandbox.
ssh 8 6 7 It's the best way to keep the agent off your own machine.
modal 9 6 6 It's a good fit for short cloud jobs and evals.
daytona 9 6 6 It's a good fit for a managed cloud workspace.
vercel_sandbox 9 5 6 It's a good fit if you already live on Vercel.
singularity 8 5 10 It's the answer for HPC clusters without Docker.

🔥 Want my exact Hermes Agent Backend Explained setup, step by step? Inside the AI Profit Boardroom you get step-by-step video trainings, my installable Agent OS and four coaching calls a week with 3,400+ members building real automations. → Join the AI Profit Boardroom

Terminal Backend Reference

local

The local backend is the default, and commands run directly on your machine with no isolation.

No setup is needed.

Commands use your real home folder, so tools such as git, ssh, gh and npm find their existing logins.

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

Hermes offsets that with a dangerous command check, which asks for approval before a risky command runs.

Verdict: It's fast and simple, and it's only as safe as the person typing.

docker

The Docker backend runs commands in a container with security hardening.

Setting What it does
docker_image It sets the base image, and the default is a Python and Node image.
docker_mount_cwd_to_workspace It mounts the folder you launched from into /workspace.
docker_volumes It mounts host folders, and you can add :ro for read-only.
docker_forward_env It forwards named secrets into the container.
docker_network It cuts the container off from the network when set to false.
container_cpu and container_memory They set the limits, and the defaults are 1 core and 5120 MB.
container_persistent It keeps one shared container when true and gives each session a fresh one when false.

Hermes starts one long-lived container and routes every call into it with docker exec.

Installed packages, files and background processes carry over from one command to the next.

The hardening flags are --cap-drop ALL, no-new-privileges and a PID limit of 256.

It needs Docker Desktop or Docker Engine, and Podman works when you set HERMES_DOCKER_BINARY=podman.

Verdict: It's the backend I'd recommend to most people once the agent does anything unattended.

ssh

The SSH backend runs commands on a remote server.

It requires TERMINAL_SSH_HOST and TERMINAL_SSH_USER, and it accepts an optional port and key path.

It keeps a persistent shell on the remote host, so the working directory and environment variables survive between commands.

The tools page recommends it for security because the agent can't modify its own code.

Verdict: It's the cleanest split between where the agent thinks and where it works.

modal

The Modal backend runs commands in a Modal cloud sandbox with configurable CPU, memory and disk.

It needs MODAL_TOKEN_ID and MODAL_TOKEN_SECRET, or a ~/.modal.toml file.

The filesystem can be snapshotted on cleanup and restored in the next session.

Verdict: It's strong isolation with no server to manage, and you pay Modal for it.

daytona

The Daytona backend runs commands in a managed Daytona workspace.

It needs DAYTONA_API_KEY, and the docs note a 10 GiB disk cap.

With persistence on, the sandbox is stopped and resumed instead of deleted.

Verdict: It's a sensible choice if you want a cloud dev box that resumes.

vercel_sandbox

The Vercel Sandbox backend runs commands in a cloud microVM.

It needs VERCEL_TOKEN, VERCEL_PROJECT_ID and VERCEL_TEAM_ID, plus an optional SDK install.

The supported runtimes are node24, node22 and python3.13.

Custom disk sizing isn't supported, and the docs say a non-default value fails instead of being ignored.

Verdict: It has the most setup steps of the seven, and it suits teams already on Vercel.

singularity

The Singularity backend runs commands in a Singularity or Apptainer container.

It needs the apptainer or singularity binary on your path.

It uses --containall --no-home, so the host home folder isn't mounted.

Verdict: It's a niche pick, and it's the right one on a shared research cluster.

🔥 Want to see my own Hermes configuration? Inside the AI Profit Boardroom, I walk through my Hermes setups on video, and you can ask me questions on four coaching calls a week alongside 3,400+ members. → Get access here

Security Comparison: Docker Versus Local

The docs publish their own security table, and I've summarised it here.

Backend Isolation Dangerous command check Docs say it's best for
local There is none, because it runs on the host. It is on. It's best for development and trusted users.
ssh It is a remote machine. It is on. It's best for running on a separate server.
docker It is a container. It is skipped, because the container is the boundary. It's best for a production gateway.
singularity It is a container. It is skipped. It's best for HPC environments.
modal It is a cloud sandbox. It is skipped. It's best for scalable cloud isolation.
daytona It is a cloud sandbox. It is skipped. It's best for persistent cloud workspaces.
vercel_sandbox It is a cloud microVM. It is skipped. It's best for cloud execution with snapshots.

The trade-off is simple.

Local protects you with a prompt, and a sandbox protects you with a wall.

A prompt only works when a human is there to read it, which is why the docs recommend a container or cloud backend for production gateway deployments.

One caution applies to Docker, because any variable you add to docker_forward_env can be read by code inside the container.

Model Backends Reference

The model backend is set with hermes model, which is the full setup wizard you run outside a chat.

/model is the quick switch inside a chat, and it only shows providers you've already configured.

Provider backend How it's set Note
Nous Portal You run hermes setup --portal or pick it in hermes model. The docs call it the recommended way to run Hermes, with 300+ models on one login.
OpenRouter You add OPENROUTER_API_KEY to ~/.hermes/.env. One key reaches many models.
Anthropic You pick it in hermes model. It supports a login or an API key.
OpenAI API You add OPENAI_API_KEY. The provider name is openai-api.
AWS Bedrock You pick "AWS Bedrock" in hermes model. It uses the standard AWS credentials chain.
Ollama You choose "Custom endpoint" and enter http://localhost:11434/v1. You should set context_length, because Ollama defaults to a small window.
LM Studio You pick "LM Studio" in hermes model. Its local server starts on port 1234.
Custom endpoint You set provider: custom with a base_url. It works with any server that implements /v1/chat/completions.

The provider table in the docs is much longer than this, and it includes Google Gemini, DeepSeek, xAI, Vertex AI and Azure AI Foundry.

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

For the AWS route, see Hermes Agent on Bedrock.

Verdict: The provider list is one of Hermes's real strengths, and the local options make a zero-cost setup possible.

🚀 Ready to put Hermes Agent Backend Explained to work in your business? Inside the AI Profit Boardroom you get step-by-step video trainings, my installable Agent OS and four coaching calls a week with 3,400+ members building real automations. → Join the AI Profit Boardroom

Web Search Backends Reference

Provider Search Extract Free tier per the docs
Firecrawl (default) It can. It can. It gives 500 credits a month.
SearXNG It can. It can't. It is free when self-hosted.
Brave Search It can. It can't. It gives 2,000 queries a month.
DuckDuckGo It can. It can't. It is free with no key.
Exa It can. It can. It gives 1,000 searches a month with a key.
Tavily It can. It can. It has a keyless option when selected.
Perplexity It can. It can. It is paid.

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

You can split the job with web.search_backend and web.extract_backend, so a free search provider can sit next to a different extract provider.

Verdict: Web search works on day one for free, which is better than most agents manage.

Browser And Memory Backends Reference

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

You pick one in hermes tools under Browser Automation, and the choice is stored as browser.cloud_provider.

I review that toolset in the Hermes Agent browser guide.

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

One external memory provider can be added with hermes memory setup, and the docs list Honcho, OpenViking, Mem0, Holographic, RetainDB, ByteRover and Supermemory.

My scored list is in the best Hermes Agent memory.

Verdict: Both layers are optional, and the defaults are good enough to start.

Hermes Agent As A Backend Reference

Mode How to start it What it serves My verdict
API server You set API_SERVER_ENABLED=true and API_SERVER_KEY, then run hermes gateway. It serves the full agent in OpenAI format on port 8642. It's the one to use for a chat frontend.
ACP server You run hermes acp. It serves editors such as VS Code, Zed and JetBrains over stdio. It's the one to use for coding.
MCP server You run hermes mcp serve. It exposes Hermes conversations to other agents over stdio. It's useful for bridging messaging platforms.
Backend server You run hermes serve. It serves the desktop app and remote clients on port 9119. It's the engine behind the desktop app.
Subscription proxy You run hermes proxy start. It serves raw model inference with no tools. It's for apps that only need the model.

The API server supports /v1/chat/completions, /v1/responses and /v1/models, and the docs name Open WebUI, LobeChat, LibreChat, NextChat and ChatBox as frontends.

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

Verdict: This is the most underrated part of Hermes, because it turns one agent into a service for every other tool you own.

Performance And Cost Notes

  • Local has no sandbox to start, so nothing sits between the agent and your shell.
  • Docker keeps one container running and attaches to it on the next launch.
  • SSH reuses one connection and one persistent shell.
  • Cloud sandboxes preserve files between sessions, and they don't preserve running processes.
  • Every backend uses a 180 second per-command timeout by default.

Hermes is free under the MIT licence, and local, Docker, SSH and Singularity add no software cost.

Modal, Daytona and Vercel Sandbox bill on their own pricing, which I haven't quoted because it changes.

🔥 Ready to build an agent stack that runs without you? The AI Profit Boardroom has my Hermes training, my agent setups and four coaching calls a week, so you can copy what works instead of guessing. → Join the AI Profit Boardroom

My Verdict

Area Score Why
Choice of backends 9/10 Seven terminal backends and a long provider list cover almost every setup.
Security options 9/10 Docker hardening, an air-gap switch and per-session containers are all built in.
Ease of switching 8/10 One hermes config set command changes the terminal backend.
Clarity 6/10 The word "backend" means six things, which confuses new users.
Overall 8/10 It's flexible and safe once you know which backend you're changing.

Related Reading

Also On Our Network

FAQ: Hermes Agent Backend

What does backend mean in Hermes Agent?

It most often means the terminal backend, which decides where shell commands execute.

The docs also use it for model, web, browser and memory providers.

What is the best Hermes Agent backend?

Local is best for getting started, and Docker is best once the agent runs unattended or takes messages from other people.

Does Hermes Agent support Docker?

Yes, and setting terminal.backend to docker runs every command in one hardened container.

Which model backends does Hermes Agent support?

It supports Nous Portal, OpenRouter, Anthropic, OpenAI, Bedrock and many more, plus local Ollama, LM Studio and any OpenAI-compatible endpoint.

Can Hermes Agent run as an API backend?

Yes, and the API server exposes it as an OpenAI-compatible endpoint on port 8642.

📺 Video notes + links to the tools 👉

🎥 Learn how I make these videos 👉

🆓 Get a FREE AI Course + Community + 1,000 AI Agents 👉

About Julian

I'm Julian Goldie, an SEO entrepreneur, author and founder of the AI Profit Boardroom, which has 3,400+ members.

I help business owners scale with AI agents, automation and SEO.

  • I run Goldie Agency, a 7-figure SEO and link building agency with 70+ people in Bangkok.
  • I've grown a YouTube channel to 400,000+ subscribers.
  • I wrote the book "Link Building Mastery".
  • I run GoldieBench, where I score AI models on real one-shot builds.

→ Get my best AI training inside the AI Profit Boardroom

Start on local, graduate to Docker, and you'll get the most out of every Hermes Agent backend.