| name | g-skl-platform-hermes |
| description | Authoritative reference for Hermes Agent (Nous Research — self-improving CLI + Telegram/Discord/Slack gateway) customization in gald3r projects. Covers ~/.hermes/ config (config.yaml/.env), AGENTS.md-native instruction loading (auto-injected with SOUL.md/CLAUDE.md/.cursorrules), agentskills.io SKILL.md skills (the gald3r distribution opportunity), delegate_task subagents, MCP (mcp_servers), the native config.yaml hooks: surface (17+ events, pre_tool_call blocking), and gald3r install verification. |
g-skl-platform-hermes
Activate for: setting up gald3r with Hermes Agent (Nous Research CLI / gateway), authoring or porting skills onto Hermes, wiring MCP via ~/.hermes/config.yaml, shipping AGENTS.md as the instruction surface, or verifying the Hermes gald3r install.
Full 9-section breakdown + evidence URLs in PLATFORM_SPEC.md (this folder). Status: ✅ near-full
parity — Hermes natively supports Skills, Rules/memory, MCP, and Hooks (plus delegate_task
subagents and an AGENTS.md-native instruction file). Commands are the only partial primitive
(skills-as-/skill-name, no command-file primitive). Headline: Hermes Skills use the
agentskills.io SKILL.md standard gald3r already ships, so gald3r skills are directly portable
— distribute them as a Hermes tap. (Re-verified 2026-06-20 against
github.com/NousResearch/hermes-agent + hermes-agent.nousresearch.com/docs — corrects an earlier pass
that wrongly marked Hooks absent.)
1. Platform Overview
Hermes Agent — Nous Research's self-improving, channel-native agent. A CLI plus a multi-platform
gateway (Telegram, Discord, Slack, WhatsApp, Signal, Email). Distinctive traits: an autonomous
learning loop that writes its own skills from experience, persistent cross-session memory, and
isolated subagents (the delegate_task model — single goal or batch tasks:[...]). All state
lives under ~/.hermes/ ($HERMES_HOME; %LOCALAPPDATA%\hermes on native Windows).
2. Config Layout
~/.hermes/ ← all state (or $HERMES_HOME)
├── config.yaml ← settings: model, toolsets, memory, compression, mcp_servers:
├── .env ← secrets only (API keys, bot tokens)
├── auth.json ← OAuth provider creds
├── SOUL.md ← persona (system-prompt slot #1, auto-loaded)
├── memories/ MEMORY.md, USER.md ← persistent memory (auto-loaded)
└── skills/ <category>/<name>/SKILL.md ← Agent Skills (agentskills.io) → /<name>
<project-root>/
└── AGENTS.md ← auto-injected into system prompt (NATIVE; also .hermes.md/CLAUDE.md/.cursorrules)
No .hermes/ project tree — config is user-global; project influence is via the auto-injected
context files only. Skills distribute as taps (hermes skills publish / hermes skills tap add owner/repo).
3. gald3r Integration
Cheapest install: ship AGENTS.md + distribute gald3r skills as a Hermes tap. Hermes auto-injects
AGENTS.md (no CLAUDE.md shim, no migration step) — fold g-rl-* rules into it. For hard
enforcement, wire g-hk-* logic into the config.yaml hooks: block: SessionStart .gald3r/
injection onto on_session_start, PreToolUse guardrails onto pre_tool_call (return
{"action":"block",…} to gate). Publish g-skl-*/SKILL.md to a repo and hermes skills tap add <repo>
— they index and surface as /<name> natively. Wire MCP via mcp_servers: in ~/.hermes/config.yaml.
Optionally add a gald3r SOUL.md persona.
Verify
Test-Path $HOME/.hermes/config.yaml # main config (mcp_servers: AND hooks: live here)
Test-Path AGENTS.md # auto-injected instruction surface
Test-Path $HOME/.hermes/skills # skills tree (g-skl-*/SKILL.md)
Test-Path $HOME/.hermes/SOUL.md # if shipping the gald3r persona
Select-String -Path $HOME/.hermes/config.yaml -Pattern '^hooks:' # confirm the gald3r hooks: block is wired
# Confirm a gald3r skill resolves: run /skills in the Hermes CLI and look for the g-skl-* entries
4. Common Pitfalls
- Hooks live in
config.yaml, not a hooks dir — g-hk-* logic wires into the hooks: block as
shell commands (JSON-stdin); pre_tool_call can hard-block via {"action":"block",…}. Caveat:
hook failures are swallowed (non-blocking-on-error), so enforce only via an explicit block action,
and verify the hook actually fires (some builds report missed firings).
- Instruction file is
AGENTS.md-native (auto-injected with SOUL.md/CLAUDE.md/.cursorrules).
Ship AGENTS.md; do not assume a .hermes/ project config tree exists.
- Skills need only
name+description — version/author/license/platforms/metadata.hermes.*
are optional/recommended; a baseline gald3r SKILL.md loads as-is (adding the three is
distribution polish, not a requirement).
- Commands = skills only — there is no user-authored command-file primitive; every skill is a
/skill-name. Distribute/update via taps, not a .hermes/commands/ dir.
- Subagents aren't declared — spawned via
delegate_task (goal/tasks:[...]), no g-agnt-*
file mapping; deliver agent behaviors as Skills.
- Gateway surfaces — in Telegram/Discord/Slack prefer summarized status (or
--json/--toon)
over wide TASKS.md markdown tables (they truncate in chat).
5. Capability Summary
| Feature | Status | Notes |
|---|
Hooks (g-hk-*.ps1) | ✅ | native config.yaml hooks: shell hooks (JSON-stdin, 17+ events incl. on_session_start/pre_tool_call); pre_tool_call blocks via {action:block}; non-blocking on hook error |
Skills (g-skl-*/SKILL.md) | ✅ | agentskills.io standard; ~/.hermes/skills/<cat>/<name>/SKILL.md; /skill-name; directly portable (only name+description required) → distribute as a tap |
Agents (g-agnt-*) | ✅ | native subagents via delegate_task (goal/batch tasks:[...]; delegation.* bounds); no declared agent file → deliver as Skills |
Commands (@g-*) | ⚠️ | built-in slash set + every skill = /skill-name; no command-file primitive |
Rules (g-rl-*) | ✅ | AGENTS.md + SOUL.md auto-injected + ~/.hermes/memories/; advisory, hardenable via a pre_tool_call hook |
| MCP | ✅ | mcp_servers: in ~/.hermes/config.yaml (stdio command/args/env OR HTTP url/headers); .[mcp] extra |
Distribution opportunity: gald3r's SKILL.md artifacts are the agentskills.io standard Hermes
uses — publish them as a tap for native one-command install onto Hermes.
Full assessment + evidence in PLATFORM_SPEC.md. Re-verify on the next @g-platform-scan-docs hermes
(crawl_max_age_days: 14). Correct docs host: hermes-agent.nousresearch.com/docs +
github.com/NousResearch/hermes-agent (the prior https://hermes.dev stub URL was wrong).
