Get The Hermes Power Cheat-Sheet

Nous Research Hermes is two things most people conflate: the open-weight Hermes models (Hermes 4, and Hermes 4.3 which pushes context to 512K) — the brain — and Hermes Agent, the open-source agent CLI/TUI with web, browser, code execution, memory, skills and MCP built in — the body. Almost everyone uses about 10% of it. This one-page (3-side) cheat-sheet is the other 90%: all 20 hidden, high-value features with the exact commands, every fact verified against Nous's own model cards and the hermes-agent docs. CLUSTER 1 — THE AGENT SETUP: one-line install (curl → hermes / hermes --tui); toolsets (web, browser, code, vision, memory, delegation, cron, skills, MCP); the Tool Gateway (hermes setup --portal — one OAuth, no API keys, provisions a model + search + image-gen + TTS + cloud browser); the skills system (browse/search/install + /slash invoke); persistent sessions (hermes --continue / -c); real terminal/code execution; runs ANY model — local (Ollama, LM Studio, vLLM) or cloud (Nous Portal, OpenRouter, Anthropic, OpenAI, Gemini); and trajectory capture (save_trajectories → ShareGPT JSONL → Nous's open-source Atropos RL framework). CLUSTER 2 — THE HYBRID BRAIN: the <think> reasoning toggle (the official deep-thinking system prompt), the API-level reasoning switch (reasoning: {enabled}), and reasoning combined with tool use (think before AND after a tool call). CLUSTER 3 — FUNCTION CALLING: the native <tools>/<tool_call> format, auto-building it with tokenizer.apply_chat_template, catching tool calls mid-stream, and the one self-hosting flag (vLLM --enable-auto-tool-choice --tool-call-parser hermes). CLUSTER 4 — STRUCTURED OUTPUT + STEERABILITY: JSON mode via <schema>, Pydantic-driven schemas with validation, OpenAI-style response_format, the counterintuitive 'omit the system prompt for the most neutral behavior' trick, and Hermes's extreme persona steerability (SOTA on RefusalBench). Plus the official Hermes 4 sampling settings (temperature 0.6 · top_p 0.95 · top_k 20), a first-24-hours checklist, and the gotchas that trip everyone up (Hermes 4 = Llama-3 format vs Hermes 3 = ChatML; the missing tool-parser flag; leaving <think> on for simple tasks; 512K is 4.3-only). Full credit to Nous Research — Hermes, Hermes Agent and Atropos are all open source.

By submitting, you'll be subscribed to the Hyperautomation AI Report newsletter — one practical AI email per week. Unsubscribe anytime.

Free. No spam. Unsubscribe anytime.