Hermes Toolbox

Public, reusable building blocks for Hermes Agent users: validated skills, the validator custom /personality preset, public validation scripts, and documented operating patterns. inventory/public-manifest.json is the source of truth for what is published; the documentation and installer are tested against it.

Repository: https://github.com/srinitude/hermes-toolbox/

Compatible Hermes release: hermes-agent 0.18.2 (release v2026.7.7.2).

Current public manifest

• Skills: 3
• Plugin packages: 0
• Profile packages: 0
• Personality presets: 1

Skills under skills/:

• hermes-agent/hermes-config-audits
• hermes-agent/honcho-memory-provider
• research/first-party-integration-audits

Personality presets under primitives/personalities/:

• validator — a validation-focused /personality overlay, installed as a config snippet and never activated implicitly.

Why the plugin and profile inventories are empty

Publication is fail-closed. A package is published only while its current source passes every provenance, privacy, completeness, real-runtime, and structural gate; a newer-but-failing candidate never replaces a last-known-good public package. The previously published tutorial plugin suite and tutorial profile package failed the hardened completeness gates themselves, so they were unpublished rather than patched in place, and no current plugin or profile candidate has passed every gate yet. Rejected candidates stay local-only and are not named here. The --plugin and --profile installer flags become usable again as soon as the manifest lists packages.

Installation

The installer is manifest-driven and dry-run by default: without --apply it prints the exact proposed destinations and writes nothing, and with no selection it lists the installable packages and exits.

git clone https://github.com/srinitude/hermes-toolbox.git "$HOME/hermes-toolbox"
cd "$HOME/hermes-toolbox"
./scripts/install-toolbox.sh

Install skills explicitly — a repeatable --skill <category/name> flag, or --all-skills for every manifest-listed skill:

./scripts/install-toolbox.sh --apply --all-skills --target "${HERMES_HOME:-$HOME/.hermes}"
./scripts/install-toolbox.sh --apply --skill hermes-agent/hermes-config-audits

Install the validator personality preset without activating it, then use it per session:

./scripts/install-toolbox.sh --apply --personalities --target "${HERMES_HOME:-$HOME/.hermes}"

/personality validator

Activation is a separate, explicit step. --activate-validator (which implies --personalities) also persists the overlay as the active default:

./scripts/install-toolbox.sh --apply --personalities --activate-validator --target "${HERMES_HOME:-$HOME/.hermes}"

--plugin <name> and --profile <name> are repeatable flags that accept only manifest-listed packages; profiles install through the real hermes profile install. Both inventories are currently empty, so the installer rejects any plugin or profile selection. There is no broad plugin flag and no enable-everything step.

See docs/install.md for the full flag reference.

Hermes Agent install prompt

Give this prompt to a Hermes Agent instance to install the toolbox for you:

Install the public Hermes Toolbox from https://github.com/srinitude/hermes-toolbox/
into `${HERMES_HOME:-$HOME/.hermes}` unless I name a different target.

1. Clone or update the repository under `$HOME/hermes-toolbox`.
2. Run `python3 scripts/validate-public-safety.py`,
   `python3 scripts/validate-identity-neutrality.py`, and
   `python3 scripts/validate-package-completeness.py`; stop on any failure.
3. Run `./scripts/install-toolbox.sh --apply --all-skills --personalities
   --target "${HERMES_HOME:-$HOME/.hermes}"`. Do not pass any plugin or
   profile selection unless `inventory/public-manifest.json` lists packages,
   and do not activate the validator personality unless I ask.
4. Prove the result with `hermes skills list --source local` and
   `hermes config check`; report commands run and any blockers.

Never read or copy credentials, memories, sessions, logs, caches, state
databases, pairing state, or runtime files.

Recommended Hermes workflow

1. Start each session with /personality validator.
2. Use /hermes-config-audits for public-safe Hermes configuration audits.
3. Use /honcho-memory-provider for current Honcho memory-provider guidance.
4. Use /first-party-integration-audits when integration claims need first-party verification.
5. Validate every result against explicit source-of-truth checks before handoff.

Publishing model

• Candidates are selected only through explicit, repeatable exporter flags or local untracked allowlists under .git/info/; there is no default sweep.
• Exports are transactional per package: a candidate is staged, sanitized, and fully validated before it atomically replaces its destination, so a failing candidate leaves the current public package byte-for-byte unchanged.
• The publisher fails closed on a dirty checkout, stages only accepted paths, and exits silently with no output when there is nothing to publish.

Validation

Before publishing or installing from a checkout, run the full packet:

PYTHONDONTWRITEBYTECODE=1 python3 -m unittest discover -s tests -v
python3 scripts/validate-public-safety.py
python3 scripts/validate-identity-neutrality.py
python3 scripts/validate-package-completeness.py
python3 scripts/validate-tutorial-suite.py
python3 scripts/verify-python-structure.py scripts tests plugins
bash -n scripts/install-toolbox.sh scripts/publish-public-candidates.sh

Safety model

This repository is intentionally public. It must not contain credentials, raw memory, sessions, logs, state databases, private profile content, private plugin state, runtime caches, pairing data, or user-specific paths. Profile and plugin packages are shareable only after sanitization, completeness, and manifest checks.

See:

• docs/install.md
• docs/recommended-hermes-workflow.md
• docs/public-safety-policy.md
• docs/identity-neutrality.md
• docs/deterministic-workflow-primitives.md
• docs/publishing-criteria.md
• docs/customization-inventory.md