Hermes Agent

Hermes Agent: Advanced

This is page 7 of 8 in the Hermes Agent guide. It covers six power features that Hermes documents: what each one is for, how to turn it on, and what you trade for it. It ends with links to the shared advanced pages.

None of this is needed for the jobs on page 5. Each feature was checked against Hermes's docs on 14 November 2026. Several of them change how much Hermes can reach, so read the trade-off before you turn one on.

Run commands in a Docker container

What it is for. The Docker backend moves every command, file change, and program Hermes runs into one long-lived container. That container is the wall that the local backend lacks. Hermes's tools page says it drops all Linux capabilities, blocks privilege escalation, and allows at most 256 processes. By default it gets 1 CPU core, 5120 MB of memory, and 51200 MB of disk.

How to turn it on. Install Docker Desktop first, and check that docker info runs in a terminal without an error. Then change the terminal: lines in config.yaml to the two lines below, and restart Hermes.

terminal:
  backend: docker

To let the container see one real folder, the configuration page uses docker_volumes. Each entry takes Docker's host_path:container_path form, and :ro on the end makes it read-only. On a Mac, the lines under terminal: could look like the example below.

  docker_volumes:
    - "/Users/YOURNAME/Documents/agent-tryout:/workspace/agent-tryout"

This guide could not confirm the right way to write a Windows folder path in that line.

You gain: a bad command lands inside the container, not on your real files. Long jobs can run without stopping for questions.
You risk: Hermes turns its approval questions off on this backend, so you can no longer approve or refuse a command before it runs. A folder you share is fully reachable unless you add :ro. Checkpoints, the next feature, also stop working.
Makes sense when: you already know Docker Desktop and want unattended jobs that touch only the container, plus at most one backed-up folder shared read-only.

Take snapshots you can roll back

What it is for. Checkpoints save a snapshot of a folder before Hermes changes files in it, so you can undo its changes with one command. A snapshot is taken before Hermes's two file-writing tools run. It is also taken before commands such as rm, mv, cp, and sed -i.

How to turn it on. Checkpoints are off by default. Type hermes chat --checkpoints to start one terminal chat with them on. To turn them on for good, add the lines below to config.yaml.

checkpoints:
  enabled: true

In a chat, /rollback lists the snapshots, and /rollback diff 1 previews what changed since snapshot 1. /rollback 1 restores it. The checkpoint page says Hermes needs Git for this and quietly turns checkpoints off without it. It also skips your whole home folder, any folder with more than 50,000 files, and any file over 10 MB.

You gain: an undo button for the agent's file changes. It keeps your own later edits unless you add --all.
You risk: a false sense of safety, because snapshots cover only the listed tools and commands. They do not cover programs Hermes writes, or anything outside the folder. They also use disk space, up to 500 MB by default.
Makes sense when: Hermes edits real files in a project folder that is not your home folder, and you still keep a normal backup.

Add tools with MCP servers

What it is for. MCP servers are add-ons that give Hermes new tools, such as GitHub, a database, or a note app. Hermes ships a catalog of servers that Nous Research staff reviewed, and every one starts off. The shared Skills and MCP page compares skills and MCP across every agent in this guide and explains how to choose safe ones.

How to turn it on. In a terminal, hermes mcp opens a picker. There, Enter installs, turns on, turns off, or removes a server. hermes mcp install deepwiki installs one catalog entry by name, and the docs describe deepwiki as a way to ask about public GitHub projects. After an install, Hermes shows a checklist of the server's tools, so you can untick any you do not want. You can also add a server by hand under mcp_servers: in config.yaml, then type /reload-mcp.

You gain: Hermes can work with services and data it cannot reach on its own.
You risk: the MCP page says installing an entry runs its setup commands and the server's own code on your computer. A server connected to an account can act in that account. Some servers also spend money on each call.
Makes sense when: you need one specific service from the reviewed catalog, and you untick every tool that writes or deletes. The account behind it should be one you could afford to lose.

Install skills from the Skills Hub

What it is for. Besides writing its own skills, Hermes can install skills other people wrote. They come from its official optional set and from public skill directories.

How to turn it on. In a terminal, hermes skills browse --source official lists the optional skills that ship with Hermes. hermes skills search followed by a word searches every source. hermes skills install followed by a skill's name installs it after a security scan. hermes skills audit scans every installed hub skill again. hermes curator lock followed by a skill's name stops the agent from changing that skill.

You gain: ready-made how-tos for jobs that would take Hermes several tries to work out.
You risk: a skill is instructions and sometimes scripts, and it shapes every later job that uses it. The skills page calls one extra scan "advisory only," so its warnings do not stop an install. Community skills also come from people outside Nous Research.
Makes sense when: the skill comes from the official set, you read its SKILL.md first, and you keep skills.write_approval: true from page 2. Hermes's skills page says that gate is off by default, so check that the line is still in config.yaml.

Run several agents with profiles

What it is for. A profile is a separate Hermes with its own settings, keys, memory, skills, chats, and schedules. You might keep a research agent apart from a phone bot, so neither one's memory leaks into the other.

How to turn it on. In a terminal, hermes profile create coder makes a profile called coder, and coder chat starts a chat with it. hermes profile list shows what exists. A blank profile copies the current model and API keys, but not bot tokens or allow lists. With --no-credentials, it copies no keys at all.

You gain: separate jobs stay separate, and each profile can have its own safety settings and Telegram bot.
You risk: every profile is another agent with your account's reach. Its own config.yaml needs the page 2 safety switches again, unless you made it with --clone, which copies the current one. The profiles page also warns never to point two running agents at one profile, because they overwrite each other's memory.
Makes sense when: you have two clearly different jobs, and you set approvals.mode: manual and cwd in each new profile before its first job.

Run a local model

What it is for. A local model keeps your prompts on your own computer and costs nothing per use. Hermes treats Ollama, LM Studio, and llama.cpp as custom endpoints that speak the OpenAI-compatible API.

How to turn it on. For Ollama, type hermes model and choose "Custom endpoint (self-hosted / VLLM / etc.)". Enter http://localhost:11434/v1, skip the API key, and enter the model's name. The block below shows the same settings as they appear in config.yaml.

model:
  default: qwen3.8:27b
  provider: custom
  base_url: http://localhost:11434/v1
  context_length: 64000

LM Studio has its own choice in hermes model, at http://localhost:1234/v1. The desktop app also has a Local Models view under Settings, then Providers, which installs and manages a llama.cpp runtime. Whichever you use, the server's own context setting must be at least 64,000 tokens. The FAQ warns that Ollama reports a model's maximum context rather than the num_ctx you set, so make the number in Hermes match.

You gain: privacy for your prompts, no per-use bill, and no rate limit from a provider.
You risk: weaker results. The memory docs say models under roughly 30B parameters often claim to save memories without saving them, and a big model needs a lot of RAM.
Makes sense when: your computer has the RAM for a model of about 30B parameters or more, and you will check its work closely.

Shared advanced topics

These shared pages go deeper on topics that apply to Hermes.

Local models in depth: how to choose a local model that fits your RAM.
Mixing models: Hermes can send side tasks, such as its background memory review, to a cheaper model.
Host access and unattended runs: Hermes's local backend is host access by default, and its schedules run with nobody watching.
An always-on machine: phone chat and schedules need the gateway running on a computer that stays on.
Subscription sign-ins: the built-in Claude Max login and Nous Portal are both account sign-ins with their own billing rules.
Relays: a relay would connect through Hermes's custom endpoint setting, model.base_url, so read this page first.

Sources

Research in this repo:

Official Hermes pages (checked 14 November 2026):

Tools and toolsets (Docker backend hardening and container defaults): https://hermes-agent.nousresearch.com/docs/user-guide/features/tools
Configuration (docker_volumes, :ro, auxiliary models): https://hermes-agent.nousresearch.com/docs/user-guide/configuration
Security (container backends skip approval questions): https://hermes-agent.nousresearch.com/docs/user-guide/security
FAQ (docker info, Ollama num_ctx): https://hermes-agent.nousresearch.com/docs/reference/faq
Checkpoints and /rollback: https://hermes-agent.nousresearch.com/docs/user-guide/checkpoints-and-rollback
MCP (catalog, hermes mcp, tool checklist, trust model): https://hermes-agent.nousresearch.com/docs/user-guide/features/mcp
Skills (Skills Hub, security scan, hermes curator lock, skills.write_approval off by default): https://hermes-agent.nousresearch.com/docs/user-guide/features/skills
Profiles: https://hermes-agent.nousresearch.com/docs/user-guide/profiles
Providers (custom endpoints, Ollama, LM Studio, 64,000-token minimum): https://hermes-agent.nousresearch.com/docs/integrations/providers
Desktop app (Local Models view): https://hermes-agent.nousresearch.com/docs/user-guide/desktop
Memory (small models and the background review): https://hermes-agent.nousresearch.com/docs/user-guide/features/memory

Next: Persona and rules.