Hermes Agent Read-Only Mode

10. června 2026

Hermes Agent Read-Only Mode

Hermes Agent Read-Only Mode

Set this in the selected profile’s config.yaml:

disabled_toolsets:

- file

- terminal

- skills

Restart Hermes and create a new session. For stronger protection, give the Hermes account Read & Execute access to the folders and use a sandbox.

Weaknesses:
the setting affects new sessions only, can be changed through another profile or account, and does not protect writes outside the specified folders.
Disabling skills also hides skill_view; write approval alone does not expose the tools. Treat configuration and ACLs as additional protection, not as a replacement for sandboxing.

On Windows, Hermes generally stores the configuration here:

%LOCALAPPDATA%\hermes\config.yaml

For example:

C:\Users\<user>\AppData\Local\hermes\config.yaml

The active profile is usually stored in the same folder, in active_profile.

Restored changes:

disabled_toolsets: [] - tools are no longer disabled.

memory.write_approval: false — memory writes need no approval.

skills.guard_agent_created: false — protection for agent-created skills is disabled.

skills.write_approval: false — skill writes need no approval.

curator.enabled: true — automatic skill maintenance is enabled.

security.tirith_fail_open: true — if Tirith is unavailable, Hermes continues.Tirith is the configured Hermes security checker for potentially risky operations.

security.allow_lazy_installs: true — automatic installation of missing components is allowed.

Three settings I would not recommend

1. `security.tirith_fail_open: true`

Tirith is the security checker. If it fails, Hermes continues anyway. Use `false` so a failed check blocks the operation.

2. `skills.write_approval: false`

Skills can be created or modified without confirmation. Use `true` to prevent unnoticed persistent changes.

3. `security.allow_lazy_installs: true`

Hermes may automatically install missing components or dependencies. Use `false` to avoid unreviewed software entering the system.

Also safer:
`skills.guard_agent_created: true` and `memory.write_approval: true`.

Žádné komentáře:

Okomentovat