Skip to content
Hussh
Connect MCP

Hermes Agent — the One runtime (Nous Research framework)

Open-source agent framework from Nous Research that Hussh forks as the runtime for One. Provider-neutral, MCP-native, skills-driven — the concrete engine behind the BYOA commitment.

Concept

TL;DR: Hermes Agent is the open-source agent framework from Nous Research that Hussh adopts as the runtime layer for One — provider-neutral, MCP-native, and skills-driven, it is the concrete engine that makes BYOA real.

Status as of 2026-05-31: see body.

Relations

  • BYOA — Hermes Agent is the runtime that delivers the Bring-Your-Own-AI/API/computer/system commitment in practice.
  • One — the personal-agent platform Hermes Agent runs underneath.
  • Kai, Nav — specialist runtimes that ride on top of the agent loop.
  • PCHP — consent protocol the agent calls through an MCP server before reading personal data.
  • Three-layer architecture — Hermes Agent sits in the Platform layer, kept model-neutral by BYOA.
  • Personal Operating Layer — Hermes Agent is the executable substrate of that layer.

Status as of 2026-05-31: Active runtime direction. Hermes Agent is a real, shipping open-source framework; the Hussh fork (hussh 🤫 One) is in active development. Treat specific deployment behaviors — which models, which MCP servers, which channels are live — as environment-specific facts that change over time.

What Hermes Agent is

Hermes Agent is a synchronous tool-calling agent framework: a conversation loop that calls a model, executes the tools the model requests, feeds results back, and repeats until the task is done. Nous Research ships it as open source, designed to be self-hosted rather than rented as a vendor service. The properties that matter to Hussh:

  • Provider-neutral. The same agent runs on any hosted, OpenAI-compatible, or cloud model API, or on local model servers (LM Studio, llama.cpp, vLLM) through a custom-provider adapter. Switching backends is configuration, not a rewrite.
  • MCP-native. A built-in MCP client connects to external tool servers at startup, discovers their tools, and exposes them as first-class agent tools. This is how the agent reaches Hussh's consent surface — by calling the consent MCP rather than touching raw personal data directly.
  • Skills-driven. Procedural knowledge lives as versioned markdown "skills" the agent loads on demand — the same raw → wiki → schema discipline Hussh uses for its own LLM Wiki, applied to agent behavior.
  • Multi-channel gateway. A messaging gateway lets one agent answer across common chat platforms (WhatsApp, Telegram, Discord, Slack, Signal, and more), so One can meet a user where they already are.
  • Persistent memory + sessions. A local, full-text-searchable session store gives durable cross-session recall, kept distinct from the consent-gated personal data vault.

Why Hussh runs Hermes Agent

Hussh's architectural bet is BYOA: the user's model and the user's data should both stay portable, with Hussh providing the consent surface, not vendor lock-in. A runtime that delivers BYOA must be (a) model-agnostic, (b) able to call an external consent layer instead of owning the data, and (c) self-hostable so the user — not an AI vendor — is the principal. Hermes Agent satisfies all three. It lets Hussh:

  • honor whatever model the user authorizes, including fully local inference for sensitive context;
  • route every personal-data read through PCHP via the consent MCP, so the agent never becomes a back door around consent;
  • ship as a fork the user or org can inspect and run themselves, keeping consent infrastructure credibly vendor-neutral.

Hermes Agent is the production-grade evolution of Hussh's earlier reference-runtime work: prior internal prototypes proved the consent-gated, local/cloud-routed, multi-channel pattern, and Hermes Agent carries that pattern forward as the actual One runtime on a maintained open-source base.

What this page is not

  • Not a claim that any specific model, channel, or integration is currently live in production — those are environment facts that change.
  • Not a second consent authority. The agent is a caller of PCHP, never a replacement for it.
  • Not a fork of Hussh's data model. Hermes Agent's own memory/session store is operational state, distinct from the consent-gated personal vault.

Sources

  • Nous Research — upstream open-source framework.
  • BYOA — the architectural commitment Hermes Agent implements.
  • One — the platform Hermes Agent powers.