Field notes · AI agents, Hermes & Claude Code By Julian Goldie

Hermes Agent Permissions Explained: Allowlists, Approval Modes, Deny Rules, MCP Trust And Approvers, Set Up Step By Step (2026)

How each permission layer works, then the exact keys to set it up with least privilege, plus the gotchas that trip teams in week one.

Get The AI Profit Stack Join AIPB →
  • 01 1,000+ done-for-you AI agent workflows
  • 02 5 live coaching calls / week with me
  • 03 7-day refund, cancel anytime
  • 04 3,400+ AI operators inside

Hermes agent permissions work in layers: who may talk to the agent at all, which shell commands pause for approval, which are refused outright, which MCP tools are loaded or must ask, and who in a chat may answer a prompt.

This page explains each layer, then sets them up in order with least privilege in mind, using only the keys in the official Hermes security reference.

The example running through every step is a small web team: one laptop running Hermes, a Slack channel where three people talk to the agent, and two nightly cron jobs.

If you'd like my team to look over your guardrail plan before you start, book a free SEO strategy session.

How The Permission Layers Fit Together

The security docs list eight layers. Four of them are permissions you set yourself.

  • User authorization. On a messaging gateway, nobody can talk to the agent unless an allowlist or a pairing code lets them in. With no allowlist set, every user is denied.
  • Dangerous-command approval. Flagged shell commands pause for a yes, or are judged by the smart reviewer.
  • Blocklist and deny rules. Some commands never run, whatever mode you pick.
  • Tool access. MCP servers can be trimmed to a list of tools, or marked untrusted so every write-capable call asks.

The rest, such as container isolation and context-file scanning, work without config.

Step 0: Let In Only The People Who Need The Agent

Least privilege starts before any command runs.

Put user IDs in ~/.hermes/.env, per platform:

SLACK_ALLOWED_USERS=U01ABC123
TELEGRAM_ALLOWED_USERS=123456789

Leave GATEWAY_ALLOW_ALL_USERS unset. The docs flag it "use with extreme caution", and without any allowlist the gateway denies everyone.

For people whose IDs you don't have yet, DM pairing works: the bot replies with an 8-character code and you approve it with hermes pairing approve <platform> <code>.

Step 0b: Know Which Commands Get Checked

Hermes compares every shell command against a curated list of dangerous patterns before it runs.

A match triggers the approval layer; anything else runs straight away.

The list lives in tools/approval.py and the docs print it in full.

It catches the obvious ones, such as rm -r, mkfs, dd if= and DROP TABLE.

It also catches quieter ones: bash -c and python -e one-liners, curl piped into sh, redirects into ~/.ssh/ or ~/.hermes/.env, systemctl restart, pkill -9, and Docker container or volume removal.

One gotcha up front: pick the docker, modal, daytona or vercel_sandbox backend and Hermes doesn't run this check at all, since it trusts the container walls instead.

So everything below applies to the local and ssh backends, where the agent works on a real machine.

Step 1: Pick The Approval Mode

Open ~/.hermes/config.yaml and find the approvals block.

approvals:
  mode: smart      # smart | manual | off
  timeout: 300     # seconds before an unanswered prompt is denied

smart is the default: an auxiliary model rates each flagged command, approves the low-risk ones for that single run, denies the clearly destructive ones and hands you the rest.

manual sends you every flagged command.

off turns approvals off entirely, which the docs reserve for CI and containers.

For the web team I'd keep smart on the laptop. The team works on live client folders, so a second opinion on every flagged command is worth the few seconds.

Step 2: Teach The Reviewer Your House Rules

Smart mode has a hook most people miss: approvals.smart_policy.

Whatever text you put there is added to the reviewer model's system prompt, through the trusted channel, never mixed with the command text.

approvals:
  smart_policy: |
    Always ESCALATE anything that touches /etc.
    APPROVE docker compose restarts inside ~/deploys.

Those two lines are the example in the configuration docs, and they fit the web team well: the deploy folder restarts often, the system folder almost never.

Keep the policy short and specific. A paragraph of vague advice gives the reviewer nothing firm to follow.

Step 3: Write A Deny List For The Never-Ever Commands

Next, list the commands the agent must never run, whatever mode it's in.

approvals:
  deny:
    - "git push --force*"
    - "dd if=* of=/dev/*"

These are fnmatch globs, matched case-insensitively against the whole command and each command inside it.

They run before --yolo, /yolo or mode: off is even consulted, which is why the docs call it yolo-with-exceptions.

Three gotchas:

  • Quote every pattern. A bare leading * is a YAML alias and the file won't parse.
  • No restart needed. The config cache is keyed on file time, so a saved change applies straight away.
  • It's a policy, not a cage. The docs say deny rules don't resolve aliases, functions or renamed binaries, so don't treat one as proof a capability is unreachable.

Underneath your list sits the built-in hardline blocklist: root wipes, fork bombs and raw disk writes are refused with no override at all.

🔥 Want the full guardrail config without writing it from scratch? 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 comparing their agent setups. → Join the AI Profit Boardroom

Step 4: Answer Prompts Without Training Yourself To Click Yes

In the CLI a flagged command looks like this:

  ⚠️  DANGEROUS COMMAND: recursive delete
      rm -rf ~/old-project
      [o]nce  |  [s]ession  |  [a]lways  |  [d]eny

Once covers this run. Session covers the pattern until you quit. Always writes it to command_allowlist forever. Plain Enter means deny.

My rule: use once by default, session for a repetitive cleanup you're watching, and always almost never by hand.

An unanswered prompt is denied after the timeout. It can't be reopened, so you just ask again, and the docs say a timeout never counts against the agent.

Step 5: Lock Down Approvals In The Slack Channel

In messaging, the agent posts the command and waits for a reply: yes, approve, ok or go to allow, no, deny or cancel to block.

The slash command reference adds /approve session, /approve always and /deny.

Here's the gotcha for the web team: by default anyone admitted to the channel can answer anyone's approval.

Fix it with one key:

approvals:
  approvers: []   # requester + listed IDs + chat admins only

With the list empty, only the person whose message raised the prompt, the scope's admins and any IDs you add can answer. A teammate who taps the button gets a private note and the prompt stays open.

Two identical commands raised by two people get two separate prompts, so one person's yes never decides the other's.

Step 6: Split Admins From Regular Users

For the slash commands themselves, add an admin list per platform.

gateway:
  platforms:
    slack:
      extra:
        allow_admin_from: ["U024BE7LH"]
        user_allowed_commands: [status]

Admins can run every slash command. Everyone else can chat normally, run /help and /whoami, and run only what you list.

The messaging docs are clear that this split gates slash commands today, not tools. More surfaces may hang off it later.

Once it's on, a non-admin can't answer an approval by any route either, because answering counts as running /approve.

Ask each teammate to type /whoami and confirm the tier they see.

Step 6b: Fence MCP Tools To What The Job Needs

Every MCP server you add brings its own tools, and by default Hermes registers all of them.

The MCP docs let you cut that down per server:

mcp_servers:
  github:
    tools:
      include: [create_issue, list_issues]
  billing:
    trust: untrusted
    tools:
      exclude: [delete_customer]

include registers only the listed tools; exclude drops the named ones. If both are set, include wins.

Per the MCP config reference, trust: untrusted makes every write-capable call on that server ask first, offering Allow Once, Session or Always. Only tools the server marks readOnlyHint: true skip the prompt.

The reference's advice is blunt: mark any server you don't fully control as untrusted.

Step 7: Decide What The Cron Jobs Do

The two nightly jobs run with nobody watching, so a prompt there would just sit until it timed out.

Hermes handles that with three keys, all defaulting to deny:

  • cron_mode for scheduled jobs.
  • single_query_mode for one-shot hermes chat -q runs.
  • unattended_mode for webhook and API server sessions.

Under deny, a flagged command is blocked and the agent has to find another path.

I'd leave all three on deny and allowlist the one rule key each job needs. Allowlisted rule keys still run under deny, so the job keeps working and nothing else slips through.

Step 8: Turn Past Approvals Into A Tidy Allowlist

After two or three weeks of real use, run:

hermes approvals suggest
hermes approvals suggest --apply 1,3

The first command reads ~/.hermes/state.db, groups the flagged commands you actually approved over the last 90 days, and prints numbered proposals with counts.

Nothing changes until you pass --apply with the numbers you pick.

Recursive deletes, sudo, disk writes, credential edits and pipe-to-shell are never proposed, however often you approved them.

To prune the list later, open hermes config edit, remove the line and restart Hermes. A running session keeps what it loaded at start.

📞 Want a second pair of eyes on your approval settings? Book a free SEO strategy session and my team will walk through how your agent, your jobs and your chats fit together. → Book your free session

Bonus Step: Gate Skill And Memory Writes Too

Shell commands aren't the only thing the agent changes. It also writes its own skills and memories.

Both have an opt-in review gate, off by default:

skills:
  write_approval: true
memory:
  write_approval: true

With the skills gate on, writes made through skill_manage are staged under ~/.hermes/pending/skills/. You review them with /skills pending, /skills diff <id>, /skills approve <id> or /skills reject <id>, from the CLI or any chat.

Memory works the same way through /memory pending, and both toggle at runtime with /skills approval on or /memory approval on.

For the web team I'd switch the skills gate on for the first month, so nobody is surprised by a new skill the agent wrote for itself.

Neither gate covers edits made directly with the file tools or the terminal.

Step 9: Use YOLO Only Where It Belongs

YOLO skips every approval prompt. It comes three ways: hermes --yolo, the /yolo toggle, or HERMES_YOLO_MODE=1.

/yolo covers just the chat you typed it in and is saved with that session, so it survives a restart. On messaging, /new and /resume bring approvals back.

A red banner and a status-bar marker stay up while it's on.

YOLO still stops at the hardline blocklist and your deny list, and an MCP server's spend limit still asks.

My use for it: a throwaway container, a scripted job I've already watched run clean, nothing else. If you want that isolation, our Hermes agent sandbox walkthrough covers the backends.

Step 10: Force Any Tool To Ask With The Community Plugin

Everything so far is about shell commands. If you want rules per tool, such as "the browser never runs this session", look at permission-tiers.

It's a community plugin by imperator-clawdius, version 0.1.0, added to the catalog in the 15 November 2026 canary. It isn't an official Nous plugin.

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

Rules go in tiers.json in the plugin's data folder: a tool-name glob, a tier of allow, ask or block, and an optional note. First match wins.

With no file, terminal* and mcp__*__* ask and every other tool keeps normal behaviour.

Its README says it logs every call to a local SQLite ledger, makes no network calls, fails closed on ask in unattended runs, and never auto-approves.

How Much Access Should You Give It?

The docs' own containment advice is OS permissions plus an isolated backend with restricted mounts, credentials and network.

In practice that means a few habits:

  • Run the agent as a user that can only reach the folders its work needs.
  • Keep a disposable container backend for experiments, and accept that it skips prompts by design.
  • Load only the MCP tools a profile uses.
  • Keep YOLO and mode: off away from machines that hold client data.

The Gotchas That Catch People

  • Container backends skip the check. Approvals don't fire on Docker or Modal at all.
  • Unquoted deny patterns. The YAML fails to load.
  • Removed allowlist entries still work. Until you restart Hermes.
  • Open approvers in a group. Anyone admitted can say yes.
  • File-tool writes are a separate layer. Protected paths such as ~/.ssh/ are hard-blocked for write_file with no prompt, while ~/.ssh/config is approval-gated.

Members of the AI Profit Boardroom trade exactly these config snippets on the weekly calls.

My Verdict On Setting Up Hermes Agent Permissions

Smart mode, a two-line policy, a short deny list and an empty approver list cover the web team completely.

Cron stays on deny with one allowlisted rule key per job.

The suggest command keeps the allowlist earned rather than guessed, and permission-tiers is worth adding only when you need per-tool rules or a call log.

If you want help shaping the policy around your own jobs, book a free SEO strategy session.

Also On Our Network

FAQ: Hermes Agent Permissions Explained

Can I force an MCP tool to ask before it runs?

Yes. Set trust: untrusted on that server and every write-capable call asks first. For any tool at all, the community permission-tiers plugin adds an ask tier.

Why didn't my deny rule load?

Most often the pattern isn't quoted. A bare leading asterisk is a YAML alias, so the file fails to parse. Quote every pattern.

Can a teammate approve a command I triggered in Slack?

Yes, by default. Set approvals.approvers to an empty list and only you, the chat admins and any IDs you list can answer.

Do approvals fire inside a Docker backend?

No. Hermes leaves container backends such as docker, modal, daytona and vercel_sandbox unprompted, relying on the container to hold any damage.

How do I let one cron job run a flagged command?

Keep cron_mode on deny and add that command's rule key to command_allowlist. Allowlisted rule keys still run under deny.

Is the permission-tiers plugin part of Hermes core?

No. It is a community plugin by imperator-clawdius that adds allow, ask and block tiers per tool name through public hooks.

📺 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

Pick the mode, write the policy, quote your deny globs, empty the approver list, and if you want the rest of the stack pre-built, join the AI Profit Boardroom or book a free SEO strategy session about your hermes agent permissions.

§ Field reports

Real wins from inside the AI Profit Boardroom

See all 3,400+ members →
AIPB member win screenshot AIPB member win screenshot AIPB member win screenshot AIPB member win screenshot AIPB member win screenshot AIPB member win screenshot AIPB member win screenshot AIPB member win screenshot AIPB member win screenshot AIPB member win screenshot AIPB member win screenshot AIPB member win screenshot

Scroll sideways →

⚑ Next step

Ready to run AI agents that actually do the work?

Join 3,400+ operators inside the AI Profit Boardroom. Get 1,000+ plug-and-play AI agent workflows, 5 coaching calls a week, and a community that holds you accountable.

Join The AI Agent Community →

7-day no-questions refund · Cancel anytime

Found this useful? Get more of this site in your Google.

One tap adds us as a preferred source — Google then shows you our posts more often in Search, Top Stories, and AI Overviews.

← Back to all posts