⭐ Get the Agent OS + join 3,400+ founders inside the AI Profit Boardroom → Join AIPB

GoldieBench Blog · 6 min read

Hermes Agent Permissions: How Allowlists, Approval Modes, Deny Rules, MCP Trust And Approvers Work, And How To Set Them Safely

Five controls decide what the agent can do without you. Here is how each one works, a least-privilege setup, and a short note on the reviewer model.

Illustrated hero: five layered gold gates in front of an agent node on a dark navy background with a faint leaderboard grid

Hermes agent permissions control how far the agent can act without you: an allowlist decides who may message it, an approval mode decides which risky shell commands pause, deny rules ban commands outright, and MCP trust settings decide which connected tools must ask first.

This page explains how each control works and how to set them for least privilege, from the official Hermes security reference.

Near the end there is one short section on the model that smart approvals depend on, with scores read on the live GoldieBench leaderboard.

If you'd rather my team reviewed your setup, book a free SEO strategy session.

Control 1: Who Can Message The Agent

On a messaging gateway, Hermes checks every sender before reading the message.

It looks at per-platform allow-all flags, the DM pairing list, platform allowlists such as DISCORD_ALLOWED_USERS, and the global GATEWAY_ALLOWED_USERS.

If none of those match, the answer is deny. A gateway with no allowlists configured turns everyone away and logs a warning.

For someone whose ID you lack, DM pairing sends them a one-time code; the owner then confirms it from the CLI using hermes pairing approve.

Control 2: Which Shell Commands Pause

Each command is compared with a curated set of dangerous patterns, such as recursive deletes, SQL DROP, writes into /etc/ and curl piped into a shell.

A match goes to approvals.mode:

  • smart, the default, has a reviewer model approve low-risk matches and refuse clearly dangerous ones, passing unclear ones to you.
  • manual passes every match to you.
  • off skips the check, as YOLO does.

On docker, modal, daytona and vercel_sandbox backends the check is not run at all, since the container contains the damage.

Control 3: Which Commands Never Run

A hardline blocklist refuses the worst commands, such as wiping the root filesystem or zeroing a disk, in every mode and with no override.

Your own approvals.deny globs sit above it and are checked before YOLO or mode: off, so "allow everything except these" is a supported setup.

The reference adds a caveat: deny globs are a command policy and do not replace OS permissions or an isolated backend.

Control 4: Which Tools Load, And Which Must Ask

Each MCP server can be cut down with tools.include or tools.exclude, or switched off with enabled: false.

The MCP config reference adds trust: untrusted, which makes every write-capable call on that server need approval. It recommends that setting for any server you don't fully control.

For built-in tools there is no core ask switch. The community plugin permission-tiers, version 0.1.0 by imperator-clawdius, adds allow, ask and block per tool name, and its README says it logs each call locally.

Control 5: Who Can Answer A Prompt

By default any admitted member of a chat can answer any pending approval.

approvals.approvers: [] narrows that to the requester, listed IDs and the chat scope's admins.

Separately, allow_admin_from splits users into admins and regular users. The messaging docs say the split gates slash commands today, not tools.

How Prompts Reach You

In the CLI, a flagged command shows four answers: once, session, always and deny, with deny as the default key.

Always writes the pattern into command_allowlist, so it never asks again. Use it sparingly.

In a chat, replies such as yes, ok, no or cancel work, as do /approve and /deny.

Hermes won't ask you to approve a command it can't show in full. If the card can't display it, only Deny is offered.

An unanswered prompt is refused after 300 seconds and can't be reopened; ask the agent again in a new message.

YOLO, And What Survives It

YOLO skips approval prompts. It starts with --yolo, the /yolo toggle or the HERMES_YOLO_MODE variable.

It still stops at the hardline blocklist and your deny globs, and three prompts still appear: an MCP server's spend limit, MCP input questions and vault card-fill prompts.

While YOLO is active, a red line at session start and a marker in the status bar make it obvious.

What Unattended Runs Do

Cron jobs, one-shot hermes chat -q runs and webhook or API sessions have nobody to answer a prompt.

Their settings, cron_mode, single_query_mode and unattended_mode, all default to deny, so a flagged command is refused and the agent looks for another way.

Rule keys in command_allowlist still run, which is how one planned job gets the single command it needs.

A Least-Privilege Setup In Six Lines Of Thought

  1. Allowlist users by ID; never set allow-all.
  2. Keep smart mode, or manual on machines with data you can't lose.
  3. Write a few quoted deny globs.
  4. Trim every MCP server and mark outside ones untrusted.
  5. Set an empty approver list in shared chats.
  6. Leave unattended modes on deny.

After a few weeks, hermes approvals suggest proposes allowlist entries from what you actually approved, skipping destructive classes, and writes nothing until --apply.

Gates On The Agent's Own Writes

Two more switches cover what the agent changes about itself rather than your system.

skills.write_approval and memory.write_approval are off by default. Switched on, they hold new skills and memories until you approve them with /skills approve or /memory approve.

The file tools also refuse to write protected paths such as ~/.ssh/, ~/.aws/ and the Hermes secret stores, with no prompt to override.

The docs call those file guards defence in depth: the terminal runs as the same user, so OS permissions remain the real boundary.

The Model Behind Smart Approvals

Smart mode is the one permission feature that depends on a model. The reviewer is set under auxiliary.approval, with provider auto and a 30-second timeout by default, which means it tries your main model.

The model configuration docs suggest a fast, low-cost model for this slot, and add that expensive models here are waste.

For the main model, the leaderboard I read on 16 November 2026 put Claude Opus 5 at 8.27, GPT-5.6 Sol at 8.16 and Gemini 3.6 Flash at 7.08 out of 10.

GoldieBench scores building tasks, not command review, so use the docs' advice for the reviewer. Tune it with approvals.smart_policy, which the configuration docs say joins its instructions.

My Verdict On Hermes Agent Permissions

The controls are layered sensibly: a closed front door, reviewed commands, a ban list that outlasts YOLO and per-server tool trust.

Set all five for least privilege, then let the suggest command keep the allowlist earned.

The model choice only matters for smart mode's reviewer, and a fast one is the documented pick.

Also On Our Network

Work through the five controls on your own install, keep the reviewer model fast, and compare notes with members of the AI Profit Boardroom or ask my team on a free strategy call while you finish your hermes agent permissions.

FAQ

What does Hermes check before running a shell command?

It compares the command with a set of dangerous patterns. A match goes through the approval mode; anything on the hardline blocklist or your deny list is refused.

Who can message a Hermes gateway by default?

Nobody. With no allowlists configured and allow-all unset, every user is denied until you add IDs or approve a pairing code.

Can I make a connected MCP tool ask first?

Yes. Mark its server trust: untrusted and every write-capable call on it needs approval.

Does any permission setting depend on the AI model?

Only smart mode. Its reviewer model rates flagged commands; the docs suggest a fast, low-cost model for that slot.

What do cron jobs do with a flagged command?

They refuse it by default, because cron_mode is deny. Allowlisted rule keys still run.

Is permission-tiers part of Hermes?

No. It is a community plugin, version 0.1.0, that adds allow, ask and block levels per tool name.

The same stack Julian uses

Run this stack yourself.

Every demo on this bench was built inside the Agent Operating System — one prompt, one shot, single HTML file out. The Agent OS, the prompts, the templates, the weekly walkthroughs and 3,400+ founders shipping with it every day all live inside the AI Profit Boardroom.

3,400+founders
258documented wins
38countries
4coaching calls a week