Aulafy
ESEN
Courses
Courses/Claude Code + Local AI/Advanced Hermes with Gemma and MLX
Lesson

Advanced Hermes with Gemma and MLX

Hermes, Gemma, and MLX form one of the most talked-about stacks for local agents on Mac. The key isn't just speed: it's designing skills, memory, limits, and tests so the agent doesn't drift.

  • Understand what MLX brings to Apple Silicon.
  • Use Hermes skills as reusable operational memory.
  • Separate fast agent, verifier, and hybrid escalation.
TerminalCode
workflow:
  planner:
    provider: "ollama"
    model: "gemma-local-fast"
    max_steps: 4
  verifier:
    provider: "local"
    model: "qwen-coder"
    can_edit: false
  escalation:
    trigger:
      - "same_error_twice"
      - "tests_fail_after_patch"
      - "missing_context"

Workflow pattern

  • Create a narrow skill for the repeated task.
  • Test it in a small repo.
  • Record decisions and frequent errors.
  • Let Hermes reuse the skill, not the full history.
  • Escalate to MoA only when a task justifies it.
TerminalCode
skill:
  name: "review-nextjs-page"
  purpose: "Revisar páginas Next.js de Aulafy"
  when_to_use:
    - "nueva lección"
    - "landing SEO"
  checks:
    - "metadata con canonical"
    - "fuentes oficiales"
    - "ChapterNav correcto"
    - "sin secretos ni datos personales"

Official sources

  • Hermes Agent documentation
  • Hermes Skills System
  • Hermes Mixture of Agents
  • MLX documentation
  • MLX LM
Local coding agent with Hermes and OllamaTool calling with local models