This is a submission for the Hermes Agent
What I Built #
Keph — an uncensored AI Guardian with a mystical sphinx persona. Users visit
keph.me, an ancient eye awakens from darkness, and they can ask anything without filters, refusals, or moral lectures.
The problem: every major AI chatbot censors itself. Users who want honest, unfiltered answers
about sensitive topics — history, philosophy, science, controversial questions — hit walls of
refusals and disclaimers. Keph removes those walls. Powered by Nous Research's Hermes model, it
delivers raw, unfiltered intelligence wrapped in a cinematic dark-gold experience.
The Guardian doesn't serve. It permits access.
Key features:
Demo #
Live site: https://keph.me Try it: Ask the Guardian anything. No restrictions. No filters.
Code #
GitHub: github.com/MAZMS/keph.me
My Tech Stack
How I Used Hermes Agent #
Hermes is the beating heart of Keph. Specifically, Hermes 4 70B
(nousresearch/hermes-4-70b
) through OpenRouter powers every single conversation — it is the
core of the product.
Why Hermes was the right fit:
Uncensored by design: Hermes is one of the few models that genuinely answers without
refusals or moral lectures. Other models claim to be "helpful" but hedge, redirect, or refuse
on sensitive topics. Hermes actually delivers. This is the entire reason Keph exists — and
Hermes is what makes it possible.
The Guardian persona: Keph's system prompt instructs the AI to embody an ancient,
all-knowing sphinx entity. Hermes follows this persona faithfully without breaking character or
injecting safety disclaimers that would shatter the immersion. It maintains the cryptic,
authoritative tone across every response.
Factual depth: When users ask hard questions — about history, science, philosophy, or
controversial topics — Hermes provides substantive, well-reasoned answers rather than
surface-level deflections. It presents all sides and lets the user decide, which aligns with
Keph's "seek the truth" philosophy.
Persona compliance: Most models will acknowledge an "uncensored" system prompt but still
refuse at the edges. Hermes actually commits to the role. It follows complex persona
instructions with genuine compliance rather than performative compliance — making Keph's
promise of "no restrictions" authentic rather than marketing.
Reliability under load: Hermes handles the full range of user queries — from casual
conversation to deeply technical questions to provocative philosophical debates — without
degrading in quality or defaulting to refusal patterns. The 3x retry logic with rate limit
handling ensures consistent delivery.
The agentic capability I leaned on most was Hermes's willingness to answer anything while
maintaining a coherent persona. Every response feels like it comes from the Guardian — ancient,
authoritative, unflinching. That's not prompt engineering alone. That's the model genuinely
playing along.