Habits

From its catalog listing

Habits gives Hermes reflexes: the tool calls and small decisions it repeats, grouped into one file of steps that runs with no model turn between the steps (the advanced llm step is the one exception); decisions go to a small typed model (Jev). "Is Hermes OK?": 41 s when Hermes works it out, 2.7 s when the reflex answers with its six checks (medians of real chats, 10 November; the fast path needs your yes and a Jev key). A reflex has four kinds of step (do, check, ask, use) and ends done, or handed back with what is left. The dream re-reads the history Hermes already keeps, with no model call, and proposes reflexes for what comes back. A reflex acts on its own only after your yes, with two exceptions named below. Try without a key: hermes habits run hermes-health.

Good to know before installing

hooks read each user message, reply and tool call; the dream reads Hermes's local session history (read only, 60 days) at most once a day (dream: manual stops it) and calls no model. Learning defaults to suggest: nothing is practised or run for you until you say try; learning: auto lets learned reflexes act with no question. A list of your reflexes (2,400 characters at most) is added to the system prompt, and Hermes may run any shipped example by name without a yes, for real (six only read; tidy-downloads moves files inside ~/Downloads). An automatic request habit can answer a message itself, without a model call, and a habit that starts on its skill runs when Hermes loads that skill. A read-only request habit you explicitly asked for starts automatic at once. Reflex commands run through Hermes's own terminal and tools under Hermes's approval rules. hermes-health sends one HEAD request to github.com and reads ~/.hermes gateway state, logs and disk use; other examples reach api.github.com, hn.algolia.com and the pages you or Hermes name. With TYPESAFE_API_KEY set, decision questions go to TypeSafe Jev (api.typesafe.ai), else to your own Hermes model. Once a message habit exists, each non-slash message of 2,000 characters or less goes there with your habit names and descriptions, and a reflex's ask steps send the text the reflex read (file names, page text). Saving a scheduled habit (every 5 minutes at most often) writes $HERMES_HOME/scripts/habits-<name>.sh and a Hermes cron job at once, before any yes, and is refused inside a cron run unless cron.allow_agent_scheduling is true; hermes habits cleanup removes every habits-* job and script. Runs and traces are stored unencrypted in $HERMES_HOME/plugin-data/habits/. No telemetry. Private Hermes imports are read-only, each with a fallback (docs/SAFETY.md §10).

From its catalog listing.

How to install

  1. Install it. Hermes asks for any API keys it needs, then asks whether to turn it on.
    hermes plugins install habits
  2. Said no? Turn it on later:
    hermes plugins enable habits
  3. Restart Hermes so it loads the plugin.

Some plugins also need a sign-in or settings: see Full setup & settings. Memory plugins ask whether to become your memory provider instead. Steps follow the official plugins guide (opens in a new tab).

Category
Automation & multi-agent
Maintained by
emiwltz (Community)
Needs
No API keys declared
Adds tools
habit_run habit_save
Adds hooks
pre_llm_call pre_tool_call post_tool_call post_llm_call post_api_request on_session_finalize transform_tool_result
Works on
linux, macos
Needs Hermes
0.22.5 or newer
Version
0.1.0
In the catalog
2026-11-12

Known issues (from its catalog listing)

  • Without a TypeSafe Jev key, steps that ask a question go to your Hermes model, slower, and those runs do not count toward your yes: a reflex with such a step stays learning.
  • Measured on one user's install, with one cheap model and one decision model (Jev); 14 of 20 request wordings never seen before were recognized, the others go back to Hermes. The headline figures are from the builds of 10 November; an earlier build of 11 November (aad995e) was tried in 229 simulated turns on throwaway Hermes profiles (docs/BENCHMARKS.md); the fixes made after it are covered by tests and by one short pass of 9 real turns on the last build.
  • The guard and the dream's exclusion of secrets are word-and-shape checks, not proofs; docs/SAFETY.md lists what they do not see.
  • With permissions: hermes, Hermes does not ask about every deletion: on Hermes 0.23 a plain rm of one file is not flagged. Set permissions: strict to make the plugin refuse deleting commands.
  • The dream proposes fixed actions only, and few real actions are local reads, so most practice counts that an action came back, not that a result matched.
  • What the dream gives on day one is a list of repeated gestures and proposals; the saving comes once reflexes are assembled or made by asking. No gain from a dream-born reflex has been measured yet.
  • The start on a loaded skill worked 2 of 2 on a real Hermes (11 November); a real MCP server: not tried.
  • A scheduled habit's cron job exists from the save, before any yes; run hermes habits cleanup before you uninstall.
Using the Hermes desktop app? See where plugins live
Hermes desktop Capabilities page, Plugins tab filtered to Official plugins built by Nous Research, each with an add button.
The official plugin catalog built into the app.Hermes desktop v0.22.5 · Capabilities > Plugins · Windows · captured Nov 1, 2026

Full setup & settings ↗Source code ↗

More in automation & multi-agent

All automation & multi-agent plugins

New Hermes Agent updates every day on X. Follow @HermesWatcher

← All plugins