Run all of this inside the full Agent OS — join the AI Profit Boardroom →

agentos.guide › blog

Hermes Agent Permissions Explained: How Approvals, Deny Rules And MCP Trust Work In Your Agent OS, Plus A Safe Setup

By Julian Goldie · 15 November 2026 · agentos.guide

Illustrated hero: gold shield and layered gates guarding a row of agent profile cards on a dark aubergine background

Hermes agent permissions decide what each profile in your Agent OS may do unsupervised: who can reach it through chat, which shell commands wait for your approval, which are banned, and which MCP tools it can load or must ask before using.

This guide explains how those layers work inside an Agent OS, sets each profile up with least privilege, and ends with three prompts that let the agent check and tighten its own settings.

Every key comes from the official Hermes security reference.

If you'd like help deciding how strict each profile should be, book a free SEO strategy session with my team.

How Permissions Work Inside An Agent OS

An Agent OS runs one gateway and several Hermes profiles, each with its own config.yaml.

That matters because permissions are set per profile. A research profile and a deploy profile can sit on the same machine with very different limits.

Four layers do the work:

  1. Who gets in. Chat users need to be on an allowlist or approved through DM pairing; with no allowlist at all, the gateway turns everyone away.
  2. Which commands pause. Shell commands matching a dangerous pattern go through approvals.mode.
  3. Which commands never run. The hardline blocklist plus your own approvals.deny globs.
  4. Which tools load. MCP servers can be trimmed or marked untrusted.

Members of the AI Profit Boardroom get the installable Agent OS, so the gateway and profiles are in place before you tune them.

The Three Approval Modes

ModeBehaviourProfile it suits
smart (default)A reviewer model approves low-risk flagged commands, refuses clearly dangerous ones, passes the rest to youResearch, writing, everyday work
manualEvery flagged command waits for youProfiles that touch live sites or client data
offNo approval checks, same as YOLOThrowaway container profiles only

A prompt nobody answers is refused after approvals.timeout, 300 seconds by default.

Profiles on docker, modal, daytona or vercel_sandbox get no command prompts at all, because Hermes relies on the container. Our Hermes agent sandbox guide explains those backends.

Is It Safe To Run A Profile Locally?

A profile on the local backend runs as your own user with no isolation, which is why the approval layer exists.

The security reference is frank that deny globs filter commands rather than wall off the system, and that the file guards are defence in depth.

For real containment it points to OS permissions and an isolated backend with restricted mounts, credentials and network.

So the safe pattern is simple: give risky profiles a container or a separate server, and keep approvals on for anything local.

Safe Setup, Step By Step

  1. Lock the front door. Put chat user IDs in ~/.hermes/.env (for example TELEGRAM_ALLOWED_USERS) and leave GATEWAY_ALLOW_ALL_USERS unset.
  2. Pick the mode per profile with hermes config edit, using the table above.
  3. Add a deny list. Three to six quoted globs; they hold even under YOLO and apply the moment you save.
  4. Bind approvals to the asker. Set approvals.approvers: [] on any profile that answers in a shared chat.
  5. Split admins. Add allow_admin_from so only you run setup slash commands; the messaging docs note this covers slash commands, not tools.
  6. Trim MCP tools. Use tools.include per server and set trust: untrusted on any server you don't fully control.
  7. Leave night jobs on deny. cron_mode already defaults to deny; allowlist only the rule key a job needs.

🔥 Want profiles that ship with sensible limits?

Inside the AI Profit Boardroom you get the installable Agent OS, step-by-step trainings, four coaching calls a week and 3,400+ members sharing their profile configs.

→ Join the AI Profit Boardroom

Making MCP Tools Ask First

An Agent OS often connects several MCP servers, and each one loads every tool it offers unless you say otherwise.

Mark a server trust: untrusted and, as the MCP config reference describes, any of its tools that can change something waits for your Allow Once, Session or Always.

Tools the server labels readOnlyHint: true stay unprompted, and the reference warns that label is only a hint from the server itself.

An unrecognised trust value is treated as untrusted, so a typo fails safe.

Forcing Built-In Tools To Ask

Hermes core offers no ask-first switch for its own built-in tools, the browser included.

The community plugin permission-tiers adds one. It appeared in the canary build dated 15 November 2026, is maintained by imperator-clawdius rather than Nous Research, and is at version 0.1.0.

hermes plugins install imperator-clawdius/hermes-permission-tiers
hermes plugins enable permission-tiers

Rules set allow, ask or block per tool-name glob. Per its README, it keeps a local call log and never auto-approves.

Three Copy-Paste Prompts For Your Agent

Paste these into the profile you're tuning. Each one shows its change before applying anything.

1. Audit this profile's permissions

Use hermes config get to read approvals.mode, approvals.deny,
approvals.approvers, approvals.cron_mode, command_allowlist,
terminal.backend and the tools and trust settings of every entry under
mcp_servers. Do not change anything. Give me a table and flag anything
looser than least privilege for what this profile does.

It uses hermes config get because the security docs make config.yaml read-denied to the file tools.

2. Draft a deny list and MCP limits

Based on this profile's job, propose 3 to 6 quoted approvals.deny
globs and, for each MCP server, a tools.include list plus a trust value.
Explain each line in one sentence. Show the YAML only. Do not edit
anything until I reply "apply".

3. Tidy the allowlist from real history

Run hermes approvals suggest and show me the numbered proposals with
their counts. Recommend which to apply and why. Do not run --apply
until I name the numbers.

The suggest command never proposes recursive deletes, sudo, disk writes or pipe-to-shell. If you remove an entry later, restart Hermes; a running session keeps what it loaded.

📞 Want your profiles reviewed by a person?

Book a free SEO strategy session and we’ll go through each profile’s access with you.

→ Book the free session

Matching Strictness To Each Profile

Most Agent OS setups end up with three kinds of profile.

If several people share one profile in a chat, sort the approver list before anything else, because by default any admitted member can answer.

Keep YOLO for disposable environments. On messaging platforms, /new and /resume switch approvals back on.

When A Prompt Arrives

On the desktop or CLI you choose once, session, always or deny; deny is the default.

In a chat, reply yes or no, or use /approve and /deny.

Prefer once. "Always" adds a permanent allowlist entry for every future session of that profile.

A Monthly Check

My Verdict On Permissions In An Agent OS

Give every profile only the people, commands and tools its job needs, and nothing else.

Allowlists, smart or manual mode, a short deny list and untrusted MCP servers cover most setups.

Use the prompts to draft the YAML, and keep the final yes for yourself.

Also On Our Network

FAQ

Can each Agent OS profile have its own permissions?

Yes. Every profile has its own config.yaml, so one can run smart mode with loose MCP access while another runs manual with untrusted servers.

How much access should an Agent OS profile get?

Only what its job needs: the users who work with it, the commands it must run and the MCP tools it uses. Everything else stays denied or asks.

Will my agent change its own permissions from these prompts?

Each prompt tells it to show the YAML first and wait. Keep that line, and apply the change yourself.

Which profiles get no command prompts?

Profiles on a container backend such as docker or modal, since Hermes treats the container as the boundary.

How do I make an MCP tool ask before it acts?

Set trust: untrusted on its server. Every write-capable call then waits for approval.

Do I need the permission-tiers plugin?

Only if a profile needs ask or block rules for built-in tools such as the browser. It is an optional community plugin.

Final word on Hermes Agent Permissions Explained

Run the audit prompt on one profile today, close whatever it flags, and if you want the whole Agent OS set up for you, the AI Profit Boardroom ships it ready to install, and a free SEO strategy session with my team can map out your hermes agent permissions.

About Julian

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

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

My YouTube channel has 400,000+ subscribers, and I run Goldie Agency, a seven-figure SEO agency.

→ Get my best AI training inside the AI Profit Boardroom

Related guides on agentos.guide

→ Hermes Vault: Save Logins, Auto-Fill 2FA And Move The Key To Your Keychain In Your Agent OS With Five Copy-Paste Prompts

→ The Best Hermes Agent Plugins: My Top 10, And Where Each One Fits In Your Agent OS

→ Hermes Agent Microsoft Teams: Set It Up In Your Agent OS

→ Hermes Agent Video Analysis: How To Set It Up In Your Agent OS

Get the whole Agent OS

The installable system behind every guide on this site — dashboard, agents, memory and pipelines — updated as new tools land, with 3,400+ members and four coaching calls a week.

Join the AI Profit Boardroom →

📺 Video notes + links to the tools 👉

🎥 Learn how I make these videos 👉

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

← all guides on agentos.guide