Hermes agent permissions are the rules that control how much access your AI agent has: who can message it, which risky actions need a person's yes, which actions are never allowed, and which outside tools it can use.
This guide explains how those rules work in plain English, then shows how much access a local shop, a service firm and a small agency should give it.
Most of it is a few lines in one config file, set out in the official Hermes security docs.
If you want to talk through your own setup with my team, book a free SEO strategy session.
Why Permissions Matter More Once Staff Use The Agent
When one person runs Hermes on their own laptop, every risky command lands in front of the same person who asked for it.
Put the agent into a shared Slack, Telegram or Teams chat and that changes.
Now several people can ask it to do things, and several people can answer its approval prompts.
Add a scheduled job that runs at night and you have a third situation: work happening with nobody there at all.
The settings differ for each of those three cases, so it pays to think about them separately.
How Hermes Permissions Work, In Plain English
Think of it as four doors.
- The front door: who can message the agent. In a chat app, Hermes only answers people on an allowlist or people you approve with a one-time pairing code. With no list set, it answers nobody.
- The checkpoint: risky commands. Commands that could delete or change important things pause for approval.
- The locked room: banned commands. A built-in list, plus your own list, that never runs.
- The toolbox: outside tools. Connected services, called MCP servers, can be limited to the tools you pick, or set so every action that changes something asks first.
Each door is a separate setting, so you can keep one tight and another relaxed.
The Control Every Install Already Has
Out of the box, Hermes checks each shell command against a list of dangerous patterns before running it.
Deleting folders recursively, wiping database tables, changing system settings, stopping services and piping a download straight into a shell all count.
The default mode is called smart. A second model reviews each flagged command: low-risk ones go ahead, clearly destructive ones are refused, and anything unclear is sent to a person.
So even a business that never touches the settings starts with a check in place.
Two other modes exist. Manual sends every flagged command to a person. Off removes the check, and the docs say to use that only in trusted, isolated environments.
Who Gets To Click Yes In A Team Chat
This is the setting most owners will want to change first.
By default, anyone admitted to the chat can answer an approval prompt, whoever raised it.
That means a junior teammate could approve a command the office manager's request triggered.
Setting approvals.approvers to an empty list fixes that. Only the person who asked, the chat's admins and anyone you name can answer. Others see a note and the prompt waits.
The messaging docs add a second control: admins, set with allow_admin_from, can run every slash command, while regular users can chat and run only the commands you allow.
Those tiers cover slash commands today. Tool access isn't split by role yet; the docs say that may come later.
Use Case: A Local Shop With One Shared Laptop
Picture a gift shop where the owner and two part-time staff use Hermes on the back-office laptop to update product sheets and draft emails.
Nobody on the team works in a terminal, so a dangerous-command prompt would be confusing.
A sensible setup here:
- Keep the mode on smart, so routine commands don't interrupt anyone.
- Add a short deny list for things that should never happen, such as force-pushing code or writing to a disk directly.
- Make sure only the owner answers approvals, if the agent is also reachable from a chat.
The deny list is the key part. It blocks matching commands even if someone later switches the approval checks off.
🔥 Want to see how other owners set their agent's limits? Inside the AI Profit Boardroom you get step-by-step video trainings, the installable Agent OS, four coaching calls a week and 3,400+ members running agents in their own businesses. → Join the AI Profit Boardroom
Use Case: A Service Firm Running Overnight Jobs
Now picture an accounting practice that schedules Hermes to tidy shared folders and build a morning summary every night.
Nobody is awake to answer a prompt, so Hermes has a setting for this: cron_mode.
It defaults to deny. A flagged command in a cron job is simply blocked, and the agent has to find another way.
The same default applies to one-shot command-line runs and to webhook or API sessions, through single_query_mode and unattended_mode.
If one specific job genuinely needs a flagged command, the safer route is to allowlist just that command's rule key rather than switch the whole mode to approve.
That keeps the night shift predictable: the job you planned runs, and anything unexpected is refused.
Use Case: A Small Agency With Client Work
An agency of five might give the agent access to client project folders and a team Slack.
Here, three settings work together:
- An empty approver list, so each person approves only their own requests.
- An admin list, so only the founders can run slash commands that change the setup.
- A deny list that blocks anything you'd never want near a client's files.
For agencies that also want a record of what the agent did, the community plugin covered below adds a per-call log.
What "Off" And YOLO Really Mean For A Business
Hermes has a shortcut called YOLO mode that skips every approval prompt for a session.
It's handy for a developer testing in a disposable environment. For a business, it mostly creates confusion about what was checked and what wasn't.
Two things soften the risk. Hermes shows a red banner and a status-bar marker while YOLO is on, and on a messaging platform starting a new conversation turns approvals back on.
Even with YOLO on, a built-in blocklist still refuses the most destructive commands, and your own deny list still applies.
My advice to owners: write in your team notes that YOLO and mode off are not used on any machine that holds client or customer data.
Keeping Prompts From Becoming Background Noise
The real risk with approvals isn't the agent. It's people clicking yes without reading.
Hermes offers two ways to keep prompts rare and meaningful.
First, the smart mode reviewer can follow your own house rules, written in plain English under approvals.smart_policy, as the configuration docs show.
Second, after a few weeks you can run hermes approvals suggest. It looks at what your team actually approved and proposes allowlist entries, without ever proposing destructive types such as recursive deletes or sudo, and changes nothing until someone applies the picks.
Fewer, better prompts are easier to take seriously.
Limiting The Outside Tools You Connect
Connecting the agent to a service such as a code host or a billing system can add many tools at once.
Hermes lets you keep only the tools you want with an include list, or drop the risky ones with an exclude list.
You can also mark a service as untrusted. Then, per the MCP config reference, every action on it that could change something asks a person first. The docs recommend this for any service you don't fully control.
For a business, that is the simplest way to make sure the agent can read your billing records without being able to change them unprompted.
The Add-On: Permission Tiers Per Tool
The built-in settings focus on shell commands and chat roles. Some owners want something simpler to explain to staff: this tool is fine, this tool asks, this tool is off.
That is what permission-tiers does. It's a community plugin, not an official Nous Research one, listed in the Hermes catalog from 15 November 2026 at version 0.1.0.
- Each rule pairs a tool-name pattern with allow, ask or block.
- With no rules set, terminal and MCP tools ask and the rest behave as normal.
- Every tool call is written to a local log you can read back with its report tool.
- Its README says it makes no network calls and stores no credentials.
It says it delivers the per-tool permission levels people asked for in issue 527. That issue, which asks for owner, admin, user and guest roles on chat platforms, is still open in the main repo.
The Risks An Owner Should Still Plan For
- Container setups skip the check. If your agent runs inside Docker or a cloud sandbox, Hermes treats the container as the boundary and doesn't prompt.
- Deny lists aren't a cage. The docs say they're a command policy, not an operating-system sandbox.
- Old allowlist entries linger. Removing one only takes full effect after Hermes restarts.
- Skills and memory are separate. If you want staff to review what the agent saves, turn on the write approval settings for skills and memory.
- A young plugin is a young plugin. Permission-tiers is version 0.1.0 from a single maintainer.
📞 Not sure which limits fit your team? Book a free SEO strategy session and we'll map out who should approve what in your business. → Claim your free session
How Much Access Should You Give It?
The short answer: only what each job needs, and no more.
- Run the agent on a computer account that can only reach the folders it works on.
- Let in only the staff who use it, by ID, rather than opening the chat to everyone.
- Connect only the services it needs, and mark the ones you don't control as untrusted.
- Keep approvals on, and never use the skip-everything mode on a machine with customer data.
The docs back this up: for real containment they point to operating-system permissions and an isolated setup with limited folders, credentials and network access.
A Simple Rollout Plan For Your Team
- Week one, owner only. Use the agent yourself on smart mode and note which prompts appear.
- Write the rules. Turn those notes into a short deny list and, if useful, a few plain-English house rules for the smart reviewer.
- Add the team. Invite staff to the chat with an empty approver list and yourself as admin.
- Add the night jobs. Keep them on deny and allowlist only what each job needs.
- Review monthly. Run the suggest command, apply what you recognise, and remove anything stale.
Rolling out in that order means every rule comes from something your business actually did, not from guesswork.
Questions To Ask Before Giving Staff Access
- Which folders and systems can the agent's machine reach, and should it reach all of them?
- Who is allowed to approve a command that changes client or customer files?
- Which slash commands, if any, should regular staff be able to run?
- What should happen when a scheduled job meets a command it isn't allowed to run?
- Do you want staff to review new skills or memories the agent saves?
Each answer maps to one setting covered above, which makes the config file easy to explain to the team.
Which Plan Or Edition Includes These Controls
There's no separate edition to buy for any of this. Approval modes, deny rules, approvers and the admin split are all part of Hermes itself and live in your own config file.
The permission-tiers plugin is installed separately from the catalog and needs Hermes 0.23 or later.
My Verdict For Business Owners
For most small businesses, the built-in settings are enough: smart mode, a short deny list, an empty approver list and cron on deny.
Add the admin split once more than two or three people use the agent in chat.
Consider permission-tiers if you need per-tool rules or a call log, and test it on one machine first.
The owners in the AI Profit Boardroom who get the most from agents tend to write their rules down before they roll the agent out to staff.
Also On Our Network
- 🌐 Every Hermes permission layer explained, with the exact keys to set each one
- 🌐 A settings table showing which Hermes permission switch fits which situation
- 🌐 How permissions work per Agent OS profile, with a least-privilege setup
- 🌐 The five Hermes permission controls explained, with a least-privilege setup
- 🌐 The complete Hermes Agent permissions guide on AI Profit Boardroom
FAQ: Hermes Agent Permissions Explained For Business Owners
Can my staff approve each other's agent commands?
By default, yes, anyone admitted to the chat can. Set approvals.approvers to an empty list so only the requester, chat admins and named people can answer.
What happens if the agent hits a risky command overnight?
Out of the box cron_mode is set to deny, so Hermes refuses that command and the job carries on another way. Nothing waits for a person.
Do I need to buy anything extra for these controls?
No. Approval modes, deny rules and approver settings are part of Hermes and live in your config file. The tiers plugin is a separate catalog install.
Can I stop regular staff from changing the agent's settings?
Yes. With allow_admin_from set, only admins run every slash command; regular users chat and run only the commands you allow.
Is YOLO mode safe for a business laptop?
It skips every approval prompt, so keep it off machines with client or customer data. The built-in blocklist and your deny list still apply.
Who makes the permission-tiers plugin?
A community developer, imperator-clawdius. It is listed in the Hermes catalog but is not an official Nous Research plugin.
📺 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've built a 7-figure SEO and link building agency, Goldie Agency, with a team of 70+ SEO experts.
- I've grown a YouTube channel to 400,000+ subscribers.
- I wrote "Link Building Mastery".
- My Udemy courses have taught over 29,000 students.
→ Get my best AI training inside the AI Profit Boardroom
Write your rules down, give approval rights to the right people, and if you want help planning it, book a free SEO strategy session or join the AI Profit Boardroom to see how other owners handle hermes agent permissions.











