Skip to content
vavio

Which actions does Hermes Agent ask me to approve?

Hermes Agent does not ask before every action. It checks shell commands for dangerous patterns, can gate an entire Python script, has a separate approval gate for edits to project instruction files, and can hold memory or skill changes for review. You can change those controls, but turning on auto-approval does not remove every gate. A remote computer can stay online while a particular action is blocked for lack of approval; hosting with Vavio does not guarantee that a job will finish unattended.

This guide follows Hermes v0.23.1, commit 3b42157, checked on 12 November 2026. We reviewed the official source and documentation; we did not run an approval trial on a Vavio computer or verify a Vavio approval interface for this article.

Which actions can ask?

Action What controls it What to expect
A shell command matching a dangerous pattern, such as recursive deletion or a service restart approvals.mode, command allowlists and the execution context A decision from smart approval, a human prompt, or a block
An execute_code Python script on a gateway or approval-enabled API surface Whole-script approval plus the execution context The script can need approval even if no shell command is flagged
A file-tool edit to a protected project instruction file security.protected_instruction_files Approval for that operation, even with auto-approval on
Saving memory or changing a skill memory.write_approval or skills.write_approval Writes proceed by default; enabling the gate can hold them for review
An MCP call crossing a configured server spend limit That MCP server’s spend_limit Separate budget approval; YOLO does not waive it

These are different decisions. A model refusing a request, a missing account connection and an operating-system permission error are separate causes of an unfinished task.

Command approval has three modes

In the profile’s config.yaml, approvals.mode can be:

  • smart, the upstream default: an auxiliary model assesses flagged commands. It can approve a low-risk command, deny a dangerous one or send an uncertain case to a person.
  • manual: flagged commands ask a person rather than using the smart decision.
  • off: skips the ordinary command approval layer. It also skips the trust prompt for write-capable tools on an MCP server marked trust: untrusted.

Manual mode does not make every command ask. Commands without a matching dangerous pattern can run, and an existing approval or allowlist entry can cover a flagged command. Execution backends also matter: some isolated backends skip ordinary command prompts; Docker with host paths mounted keeps those checks.

When a command prompt offers once, session or always, those choices have different reach. Once authorizes that execution. Session covers its detected pattern for the session. Always saves a permanent allowlist entry in command_allowlist for future sessions. A broad pattern can cover more than the command you first approved. Where you have the Hermes command line, hermes config edit lets you review those entries. After removing an entry, restart Hermes so the running process does not retain an already-loaded approval.

What happens when nobody can answer?

Cron jobs, one-shot queries and unattended webhook/API sessions have separate settings. These upstream defaults deny an action that reaches their approval gate:

approvals:
  cron_mode: deny
  single_query_mode: deny
  unattended_mode: deny

single_query_mode covers hermes chat -q. unattended_mode covers programmatic platforms without a human approval route; an API client that can answer approval cards can still receive a prompt. A messaging session with an approval channel may wait for a reply instead. The upstream default timeout is 300 seconds; an unanswered prompt denies the action.

Setting one of these keys to approve allows actions handled by that context’s ordinary gate. It does not authorize every operation in the system. Permanent allowlist entries, YOLO or mode: off can bypass an ordinary gate before its unattended deny setting is considered. Conversely, a separate protected-file or budget gate can still block an action.

A session-only shell approval does not override unattended deny; that path checks permanent approvals. Smart mode also does not assess a flagged shell command before this unattended decision.

execute_code needs special attention: its Python can start processes or change files without passing through shell-pattern checks. Gateway and relevant API/automation turns therefore gate the whole script; unattended deny can block it even when the script looks harmless. Local interactive CLI turns, including viewer-submitted local TUI/Desktop turns with an approver attached, normally rely on per-call terminal checks instead of a whole-script prompt. Isolated backends can skip this gate.

A blocked command is returned to Hermes as a tool result. That does not prove the whole job stopped, found another solution or completed correctly. Read the job’s output and the actual result before treating it as finished.

What does YOLO leave in place?

/yolo toggles auto-approval for the current chat. A CLI launch with --yolo or HERMES_YOLO_MODE=1 applies to that process’s sessions; approvals.mode: off applies across the profile. None is a promise of unrestricted execution.

The normal command path retains hardline blocks and user-defined approvals.deny rules. Supervised gateways also guard against stopping or restarting their own process. MCP spend-limit approvals, requests for missing input, payment-card fill prompts and MCP event-subscription prompts are separate. The protected project instruction-file gate below also remains.

These checks are application controls. They do not replace operating-system permissions or isolation, and file-tool guards do not contain every possible shell action.

Project instructions ask separately

By default, write_file and patch ask before changing project files named AGENTS.md, CLAUDE.md, SOUL.md or .cursorrules, ignoring capitalisation. Files directly inside a project-local .hermes directory are also covered. Each operation needs its own approval; session or permanent command approval does not cover it. Without a working human channel, the write fails rather than auto-approving.

The gate is controlled by security.protected_instruction_files (default true); security.protected_instruction_extra_patterns adds filename patterns. Names such as AGENTS.override.md and .hermes.md need an added pattern to receive this protection. Hermes’s own profile home and global home are exempt from this project-file check, so it does not stop edits to the profile’s own SOUL.md. See which instructions Hermes loads before choosing a file.

Memory and skills can wait for review

memory.write_approval and skills.write_approval default to false, allowing those subsystem writes. Setting either to true enables its own gate, independent of command auto-approval.

With the memory gate on, foreground memory writes can ask inline through an interactive CLI; without that prompt route, they are staged for review. Skill writes and background-review writes are staged. A staged change has not been committed to the memory or skill it would change. In a supported interactive session, /memory pending or /skills pending shows pending work. Headless runs point to pending records under the profile home for later review. This distinction can explain why a new chat does not remember a proposed fact.

Choose autonomy for the job you want

Before leaving a recurring task alone, identify its tools, account access, approval settings and observable output. Try the same work with those settings and check the result. A successful chat reply alone does not qualify an unattended workflow.

Vavio provides a managed Hermes computer for $14.99 per month, with model usage billed separately. The included MCP server and CLI let a supported assistant help manage that setup. Hosting supplies a computer away from your laptop; the approval rules still depend on the Hermes version, profile and client in use. If that is the service you need, compare the current hosting offer. If you already self-host, check the same settings in your own profile before changing them.

Sources

Written by Vavio with AI assistance and checked against pinned official Hermes v0.23.1 source and documentation. This is a source review, not a live approval trial.

More from the Vavio blog: Hermes Agent guides