Hermes agent permissions are a set of separate switches, and the right one depends on what you want to stop: strangers messaging the bot, risky shell commands, a specific tool, a connected MCP service, or a teammate approving something they shouldn't.
This page explains what each switch does and when to reach for it, then lists the mistakes that leave an agent with more access than it needs.
Everything here comes from the Hermes security reference, the MCP config reference and the permission-tiers plugin's own repository.
If you want my team to help you choose, book a free SEO strategy session.
Does Hermes Agent Have Restrictions Out Of The Box?
Yes, several, before you change anything.
- A messaging gateway with no allowlist denies every user.
approvals.modestarts onsmart, so flagged shell commands are reviewed rather than run blind.- A hardline blocklist refuses root wipes, fork bombs and raw disk writes in every mode.
- Unanswered prompts are denied after 300 seconds.
- Cron jobs, one-shot
hermes chat -qruns and webhook sessions deny flagged commands. - File tools can't write protected paths such as
~/.ssh/or the Hermes secret stores.
What isn't restricted by default: every tool an MCP server offers is loaded, every admitted chat member can answer approvals, and every allowed chat user can run every slash command.
Every Permission Setting, Side By Side
| Setting | What it does | When to use it |
|---|---|---|
Platform allowlists, e.g. SLACK_ALLOWED_USERS | Decides who may message the agent at all | Always, on any messaging gateway |
| DM pairing | Unknown users get a code you approve with hermes pairing approve | When you don't have someone's ID yet |
approvals.mode: smart | A reviewer model approves low-risk flagged commands, denies dangerous ones, asks you about the rest | Most interactive use |
approvals.mode: manual | Every flagged command waits for a person | Machines with data you can't lose |
approvals.mode: off / YOLO | No approval prompts | Disposable containers and CI only |
approvals.deny | Glob list that blocks commands even under YOLO | The handful of commands that must never run |
command_allowlist | Flagged patterns that run without asking | Routine commands you have approved many times |
approvals.approvers | Who may answer a prompt in a chat | Any group or shared channel |
allow_admin_from | Splits admins from regular users for slash commands | Once more than one person uses the bot |
MCP tools.include / exclude | Loads only chosen tools from a server | Any server with tools you don't need |
MCP trust: untrusted | Every write-capable call on that server asks | Any server you don't fully control |
| permission-tiers plugin | Allow, ask or block per tool name, plus a call log | When you need a rule for a non-MCP tool |
Choosing An Approval Mode
The mode only matters on backends where the check runs. On docker, modal, daytona and vercel_sandbox, Hermes skips dangerous-command prompts and treats the container as the wall.
On local and ssh, pick by how much a wrong command would cost you.
- Smart suits daily work. Tune it with
approvals.smart_policy, plain-text rules the configuration docs say are added to the reviewer's instructions. - Manual suits a server that hosts client work, where you want to read every flagged command.
- Off suits nothing that holds real data. The docs limit it to trusted environments such as CI and containers.
If smart keeps refusing variations of the same command, approvals.denial_breaker_threshold, default 3, makes the agent stop and ask you.
Choosing Between Deny Rules And The Allowlist
These pull in opposite directions, and people mix them up.
approvals.deny is for "never". Its quoted globs are checked before YOLO, /yolo and mode: off, so a deny rule holds even when approvals are switched off.
command_allowlist is for "always fine". It fills up when you answer "always" to a prompt, or when you apply proposals from hermes approvals suggest.
The suggest command never proposes recursive deletes, sudo, disk writes, credential edits or pipe-to-shell, which makes it the safer way to grow the list.
๐ฅ Want a pre-built agent with these switches already set? 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 sharing their configs. โ Join the AI Profit Boardroom
Choosing Who Can Approve In A Chat
Two settings answer two different questions.
approvals.approvers answers "who can say yes to this prompt?" Set it to an empty list and only the requester, any IDs you add and the chat scope's admins can answer.
allow_admin_from answers "who can run setup commands?" Admins run every slash command; regular users chat and run only what user_allowed_commands lists, plus /help and /whoami.
If both are set, a narrowed approver list takes precedence for answering prompts, per the messaging docs.
The admin split gates slash commands today, not tools. The docs mention tool access as something that may be gated later.
Choosing How To Limit MCP Tools
An MCP server can expose a few tools or thousands. The MCP docs give three levers.
- Turn the server off with
enabled: falsewhen you don't need it at all. - Trim it with
tools.includefor a short list, ortools.excludeto drop a few dangerous ones. Globs work, andincludewins if both are set. - Make it ask with
trust: untrusted. Every write-capable call then needs approval, with Allow Once, Session or Always.
The config reference says readOnlyHint is only a hint supplied by the server, so mark any server you don't fully control as untrusted.
A declared spend_limit still asks even under YOLO.
Choosing How To Force Any Tool To Ask
Want the browser or file writes to ask every time? Core Hermes can't do that per tool.
Permission-tiers adds one. Its maintainer is imperator-clawdius, its version is 0.1.0, and the plugin catalog files it under community, not official or verified.
- Rules pair a tool-name glob with
allow,askorblock; the first match wins. - With no rules,
terminal*and MCP tools ask and everything else is untouched. - Its README says it logs every call to a local SQLite file and makes no network calls.
The README presents it as an answer to issue 527, a still-open request for role-based access on chat platforms.
Is It Safe To Run Hermes Agent Locally?
On the local backend the agent works directly on your machine, as your user, with no isolation.
The approval layer and blocklist still run there, which is the point of keeping them on.
The docs are careful about the limits. Deny rules are a shell-command policy, not a sandbox, and the file-write guards are defence in depth, since the terminal runs as the same user.
For containment they recommend operating-system permissions and an isolated backend with restricted mounts, credentials and network. Our Hermes agent sandbox guide compares those backends.
How The Prompts Behave Once Set
Knowing what an approval looks like helps you pick the right mode.
- CLI: once, session, always or deny. Deny is the default key, so a stray Enter blocks.
- Chat: a short yes or no reply works, as does typing
/deny./approve sessionand/approve alwayswiden the answer. - Long commands: Hermes sends the full text first; if it can't, only Deny is offered.
- Timeouts: no answer in 300 seconds means deny, and an expired prompt can't be reopened.
- MCP trust prompts: Allow Once, Session or Always, each covering one tool, never the whole server.
What YOLO Does And Doesn't Override
YOLO comes three ways: --yolo, the /yolo toggle and HERMES_YOLO_MODE.
It overrides approval prompts and the untrusted-MCP prompt.
It does not override the hardline blocklist, your approvals.deny globs, an MCP server's spend limit, MCP input questions or vault card-fill prompts.
/yolo covers only the chat it was typed in; approvals.mode: off covers every chat and cron job of that profile. That difference decides which one, if any, you should ever use.
Skill And Memory Writes: A Separate Switch
None of the settings above cover what the agent saves about itself.
skills.write_approval and memory.write_approval, both off by default, stage those writes until you approve them with /skills approve or /memory approve.
Turn them on when you want to review new skills before the agent relies on them.
Common Mistakes
- Setting
GATEWAY_ALLOW_ALL_USERS. The docs warn "use with extreme caution"; use allowlists instead. - Leaving approvers open in a group. Any admitted member can approve anyone's command.
- Unquoted deny globs. A bare leading asterisk breaks the YAML.
- Trusting every MCP server. The default trust is full.
- Answering "always" by reflex. It writes a permanent allowlist entry.
- Removing an allowlist entry without restarting. A running session keeps what it loaded.
- Assuming Docker prompts. Container backends skip the check by design.
- Treating permission-tiers as official. It's community code at 0.1.0.
๐ Not sure which switches your setup needs? Book a free SEO strategy session and my team will go through your agent's access with you. โ Book the free call
Choosing What Scheduled Jobs May Do
Three switches govern sessions with nobody watching: cron_mode for cron jobs, single_query_mode for one-shot runs and unattended_mode for webhook and API server sessions.
Each accepts deny or approve, and each ships as deny.
Switching one to approve lets every flagged command through in that context, which is rarely what you want.
The narrower choice is to keep deny and add the specific rule key a job needs to command_allowlist. Rule keys are honoured on unattended surfaces too, so the planned command runs and nothing else does.
Who Each Setup Suits
- Solo user on a laptop: smart mode, a short deny list, MCP servers trimmed to what you use.
- Small team in Slack or Telegram: add allowlists by ID, an empty approver list and an admin list.
- Anyone with scheduled jobs: leave cron, single-query and unattended modes on deny; allowlist one rule key per job.
- Teams that need per-tool rules or an audit log: trial permission-tiers on one machine.
My Verdict On Which Settings To Use First
Start with the front door: allowlists by ID, never allow-all.
Then keep smart mode, write three or four deny globs, and mark outside MCP servers untrusted.
Add the approver and admin settings the day a second person joins the chat.
Members of the AI Profit Boardroom swap these exact configs on the weekly calls.
Also On Our Network
- ๐ Every Hermes permission layer explained, with the exact keys to set each one
- ๐ A plain-English explainer on how much access a business should give the agent
- ๐ How permissions work per Agent OS profile, with a least-privilege setup
- ๐ The five Hermes permission controls explained, with a least-privilege setup
- ๐ The complete Hermes Agent permissions guide on AI Profit Boardroom
FAQ: Hermes Agent Permissions Explained
Does Hermes agent have restrictions by default?
Yes. A gateway with no allowlist denies everyone, smart approvals are on, the hardline blocklist always applies, and cron jobs deny flagged commands.
Which setting stops a command even in YOLO mode?
approvals.deny. Its glob patterns are checked before YOLO and mode off, and the built-in blocklist sits beneath them.
How do I make an MCP server ask before changing anything?
Set trust: untrusted on it. Every write-capable call then needs approval, with Allow Once, Session or Always.
What is the difference between approvers and admins?
approvals.approvers decides who can answer a prompt. allow_admin_from decides who can run every slash command.
Can I make a built-in tool like the browser ask first?
Not in core Hermes. The community permission-tiers plugin adds allow, ask or block per tool name.
Is it safe to run Hermes on my own computer?
The approval layer and blocklist run on the local backend, but there is no isolation. For containment the docs point to OS permissions and an isolated backend.
๐บ Video notes + links to the tools ๐
๐ฅ Learn how I make these videos ๐
๐ Get a FREE AI Course + Community + 1,000 AI Agents ๐
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.
โ Get my best AI training inside the AI Profit Boardroom
Pick the switches that match your situation from the table, fix any mistake on the list, and if you want a hand, talk to my team on a free SEO strategy session or bring your config to the AI Profit Boardroom to finish your hermes agent permissions.











