Skip to content

pi-agent-setup.md

Latest commit

 

History

History
86 lines (63 loc) · 1.71 KB

pi-agent-setup.md

File metadata and controls

86 lines (63 loc) · 1.71 KB

Pi Agent Setup

Use this guide to run Personal OS with Pi as the agent frontend.

Optional: Fully Local Backend on Mac

If you want fully local/offline execution, run a local model server first.

Local server with llama.cpp

brew install llama.cpp
llama-server \
  -hf unsloth/GLM-4.7-Flash-GGUF:UD-Q4_K_XL \
  --jinja \
  --temp 0.7 --top-p 1.0 --min-p 0.01 \
  --repeat-penalty 1.0 --fit on \
  --port 8080

Quantization guidance:

  • 16 GB RAM: UD-Q3_K_XL
  • 24 GB RAM: UD-Q4_K_XL
  • 36 GB+ RAM: UD-Q6_K

Fallback model for smaller machines:

llama-server -hf bartowski/Qwen2.5-Coder-7B-Instruct-GGUF:Q4_K_M --fit on --jinja --port 8080

Connect Pi to local server

npm install -g @mariozechner/pi-coding-agent

Create ~/.pi/agent/models.json:

{
  "providers": {
    "llama-cpp": {
      "baseUrl": "http://localhost:8080/v1",
      "api": "openai-completions",
      "apiKey": "none",
      "models": [{ "id": "GLM-4.7-Flash" }]
    }
  }
}

Run pi, select your local model, then continue with repo setup below.

What Pi Needs

  • This workspace as working directory
  • Access to AGENTS.md (shared behavior)
  • Access to canonical skills in .agents/skills/

Recommended Setup

  1. Open Pi in:
<repo-root>
  1. Ensure Pi uses this repo's canonical skills path:
<repo-root>/.agents/skills
  1. Start with:
Read AGENTS.md and run the daily standup workflow

Optional: Local-Only Runtime

If you run Pi with a local model backend, this repo still works unchanged. Skills and workflows are model-agnostic.

Notes

  • PI.md and AGENTS.md are wrappers/instructions only.
  • The source of truth for skills remains .agents/skills/*/SKILL.md.