/specialists / local-llm-tinkerer

Local LLM Tinkerer

Specialist agent for serving, benchmarking, quantizing, and evaluating local LLMs safely on your own hardware, built on the Agent Wikis Pro local-LLM skills.

v1.2.0 · 9 skills · 3 tools · 3 routines

What it ships with

Skills, loaded automatically when a task calls for them:

• local-llm-benchmarking
• local-llm-evaluation
• local-model-quantization
• benchmark-audit
• llm-inference
• llm-training
• local-model-hermes-profiles
• dgx-spark-ops
• background-process-cleanup

Tools:

• local-bench — local-bench is a standalone, local-model-only deployment qualification suite for OpenAI-compatible endpoints. It has no runtime dependencies of its own and never starts, stops, or downloads a model.
• memory-watchdog — A small guard that runs beside one model-serving unit and stops that unit before a unified-memory host (DGX Spark / GB10 and similar) runs out of memory. It never stops anything other than the unit you name.
• mptq — Shrinks an expert-heavy mixture-of-experts model into a small GGUF by keeping its sensitive layers at high precision and compressing the routed experts only where an importance matrix says it's cheap.
• Launch templates — safe starting points for serving models, with a readiness check and a memory guard.

Routines, scheduled workflows that ship switched off:

• campaign-status — every 30 min
• campaign-deadline-guard — every 5 min, no model needed
• nightly-regression — daily at 03:00

Install

# import the download directly
hermes profile import local-llm-tinkerer-1.2.0.tar.gz

# or unpack and install (makes later updates one command)
tar -xzf local-llm-tinkerer-1.2.0.tar.gz -C ~/agent-dists
hermes profile install ~/agent-dists/local-llm-tinkerer --alias

hermes -p local-llm-tinkerer model      # pick the agent's own model
local-llm-tinkerer chat

Your memories, conversations, keys, and settings stay yours: updates replace only what the specialist ships.