Hermes Agent Integration

Live on the operator's node at ~/.hermes/profiles/frameserv2/ — verified end-to-end ("ARISE" round-trip through hermes → tailnet → llama-swap → MTP inference):

~/.hermes/profiles/frameserv2/config.yaml
---
model:
  default: "qwen38-27b"
  provider: "custom:frameserv2"
  base_url: "http://100.126.153.18:8081/v1"
  api_key: ""
providers:
  custom:frameserv2:
    base_url: "http://100.126.153.18:8081/v1"
    model_prefix: ""
    api_key: ""

agent:
  max_turns: 60
  task_completion_guidance: true

display:
  personality: technical
  show_cost: false
  show_reasoning: false
  compact: true

Run it: hermes --profile frameserv2 · one-shot: hermes --profile frameserv2 -z "…"

Switch models by changing model.default to any of the model ids above — one endpoint, three brains:

• qwen38-27b — default. 19.3 t/s served, MTP-accelerated, full vision + tools. Fastest all-rounder.
• flashnext-iq1s — 176B MoE at 31 t/s decode. The heavy hitter.
• glm53-flash — 313B MoE at 14.1 t/s. Deep reasoning; structured tool calls work (template fix shipped on the box 2026-09-10).

Multiple profiles can share this endpoint with different personas (SOUL.md) — same pattern as zaraki/ishigami on ashborn-prime.

from openai import OpenAI
client = OpenAI(base_url="http://100.126.153.18:8081/v1", api_key="…")
r = client.chat.completions.create(
    model="qwen38-27b",
    messages=[{"role": "user", "content": "Hello"}],
)
print(r.choices[0].message.content)

Security posture: API-key auth is enforced on every inference endpoint (401 without a valid key; two independent keys live since 2026-09-12) on top of tailnet-scoped reachability. Verified: 401 with a wrong or missing key, 200 with either key.

The Model Playground (on the box, tailnet-only) runs one prompt against every model sequentially with live streaming — the fastest way to see the differences.