skills.rest
SouthpawIN
SouthpawINCOMMUNITY
Explore all skills by this author ā†’
GitHub

turbofit

Configures and operates an adaptive local inference provider for Hermes Agent.

95|22|Updated Jun 22, 2026
Search Tags:#tailscale#openai-compatible#llama.cpp#local inference#hermes agent#model quantization#hardware detection
One-click install
npx skills add https://github.com/SouthpawIN/turbofit --skill turbofit-southpawin
Or copy as Structured Prompt for Agentā–¼
Please help me install this Agent Skill.
Skill: turbofit
Source: https://github.com/SouthpawIN/turbofit
Command: npx skills add https://github.com/SouthpawIN/turbofit --skill turbofit-southpawin

SYSTEM DOCUMENTATION & REQUIREMENTS

šŸ’” This Skill requires pyyaml, and includes scripts (resource) and references (resource) and assets (resource) components.

What problem does it solve? Choosing and running the right local LLM for a given machine is error-prone: VRAM versus total RAM, quantization choices, context windows, and engine compatibility all interact. This Skill inspects the hardware, applies an evidence-backed model ladder, and exposes one stable OpenAI-compatible endpoint so Hermes Agent always talks to custom:turbofit / auto regardless of what runs underneath. ## Core Features & Use Cases - Hardware-aware configuration: turbofit_status and turbofit_configure probe dedicated VRAM, unified memory, or RAM-only hosts and select Fit List profiles (Maple at 8 GB, Unleashed at 16ā€“95 GB, Qwen 3.8 16-bit at 96 GB+). - Adaptive runtime ladder: /turbofit shift up|down|<model>|intelligence|balanced|speed walks the measured ladder, contracting context, auxiliary mode, and model size under memory pressure and healing back up. - Provider and fallback management: registers the named turbofit provider, sets it as primary or appends it to the fallback chain, and publishes the :8091 gateway privately over Tailscale Serve. - Use Case: On a 16 GB GPU workstation, run /turbofit scan to get the Unleashed UD-IQ3_XXS recommendation, then /turbofit setup to download the pinned artifacts and start the local stack, and verify with curl http://127.0.0.1:8091/v1/models. ## Quick Start Ask the agent to run /turbofit status to inspect the current provider, gateway health, and active rung, then run /turbofit setup to download the recommended models and start the local runtime.

Frequently Asked Questions about turbofit

High-intent search queries and answers about installing and using this skill.

FAQPage Schema
How do I set up a local LLM provider for Hermes Agent?ā–¼

Install the Turbofit plugin with hermes plugins install --enable, then run /turbofit setup. Setup downloads the recommended models for your hardware, starts the local stack on 127.0.0.1:8091, and registers the custom:turbofit provider with model auto.

How do I choose the right local model for my GPU memory?ā–¼

Run /turbofit scan to inventory dedicated VRAM, unified memory, or RAM-only capacity. The Fit List maps 8 GB to Maple Preview TQ2_0, 16 GB to Unleashed UD-IQ3_XXS, 24ā€“95 GB to UD-Q3_K_XL, and 96 GB+ to Qwen 3.8 16-bit.

Does Turbofit work on Apple Silicon Macs?ā–¼

Yes, Apple Silicon uses MLX with OrcaRouter Uncensored 4/6/8-bit quants, never 2-bit. Macs under 24 GB unified memory stay on Ornith 1.5 35A3B, and Maple uses the deepgrove MLX build.

Why does /turbofit return connection refused on port 8091?ā–¼

A refused connection to 127.0.0.1:8091/v1/models means the Turbofit runtime is not running, not a firewall or Hermes messaging-gateway issue. Run /turbofit setup or check turbofit_status from Desktop or Sirvir to restart the stack.

Can I share my local model server with other devices?ā–¼

Yes, /turbofit serve publishes the 8091 gateway over Tailscale Serve as a private HTTPS endpoint on your tailnet. Other devices use that URL as their OpenAI base URL; Funnel is never used and public binds are rejected.

What inference engines does Turbofit support?ā–¼

Turbofit auditions llama.cpp, MLX, SGLang, vLLM, FreeToken, and Turbohaul Manager against the selected model pair. Maple GGUF requires the Maple llama.cpp fork, while vLLM and SGLang serve Qwen HF/FP8/NVFP4 recipes.

Related Skills

Explore More Software Engineering Skillsā†’
GeargrindadminGeargrindadmin

llmfit-advisor

Analyze local CPU, RAM, and GPU/VRAM to recommend LLM models with quantization and fit scores.

Community
Intermediate
Supporter09Supporter09

speculative-decoding

Accelerate LLM token generation with speculative decoding strategies.

Community
Advanced
mr-kellymr-kelly
4

kelly-creators

Manage influencer marketing workflows across discovery, planning, activation, and measurement phases.

Community
Advanced
OpenRouterTeamOpenRouterTeam
209

openrouter-models

Retrieve real-time metadata for 300+ OpenRouter AI models.

Official
Intermediate