Operate Opt-In Built-In Hermes Plugins

Hermes includes a small set of in-tree plugins, but bundled does not mean automatically enabled. Most are opt-in because hooks and tools can change runtime behavior, create external telemetry, or add new authority.

Official documentation: https://hermes-agent.nousresearch.com/docs/user-guide/features/built-in-plugins

Understand discovery and precedence

Hermes discovers bundled, user, project, and package plugins in a defined order. Later sources can override an earlier plugin with the same name. Project plugins should remain disabled unless the environment intentionally allows repository-provided code to load.

Review hooks as production code

A hook can observe or alter file operations, model calls, session lifecycle, and external telemetry. Read the plugin description, dependencies, data destination, state files, and failure behavior before enabling it. Start in a disposable profile when the plugin touches cleanup or security-sensitive paths.

Operational checklist

  • Run hermes plugins list.
  • Read the specific plugin documentation.
  • Identify hooks, tools, commands, and external endpoints.
  • Enable one plugin at a time.
  • Restart when the plugin lifecycle requires it.
  • Test success, failure, and disable paths.
  • Document state files and cleanup behavior.

Common pitfalls

  • Assuming bundled plugins are already active.
  • Ignoring name collisions with user or project plugins.
  • Sending observability data to an external service without review.
  • Enabling cleanup behavior without a dry run.
  • Treating warning-only security guidance as enforcement.

Verification steps

  1. Confirm the plugin appears in the list.
  2. Enable it explicitly and restart the relevant process.
  3. Exercise one harmless documented behavior.
  4. Inspect logs or state without exposing sensitive data.
  5. Disable the plugin and confirm the behavior stops.
  6. Verify unrelated plugins and core tools remain unaffected.

A reliable Hermes workflow is defined by observable behavior, bounded authority, and repeatable verification—not by configuration alone.