Configuration for pi and omp coding agents backed by a local oMLX inference server on Apple Silicon, or koboldcpp on WSL.
See docs/setup.md for full installation instructions on macOS and Ubuntu/WSL.
# 1. Clone and symlink
git clone https://github.com/pythoninthegrass/pi_config.git ~/git/pi_config
cd ~/git/pi_config
mkdir -p ~/.pi/agent
ln -sf $(pwd)/models.json ~/.pi/agent/models.json
ln -sf $(pwd)/settings.json ~/.pi/agent/settings.json
ln -sf $(pwd)/.mcp.json ~/.pi/agent/.mcp.json
ln -sf $(pwd)/themes ~/.pi/agent/themes
ln -sf $(pwd)/extensions ~/.pi/agent/extensions
# 2. Configure .env (see .env.example)
cp .env.example .env && $EDITOR .env
# 3. Render settings and launch
source ~/git/bashrc/.bashrc && pi# Verify oMLX is reachable and model is loaded
source .env
curl -s http://127.0.0.1:8000/v1/models \
-H "Authorization: Bearer $OMLX_API_KEY" | python3 -m json.tool
# Benchmark pi
time pi -p "Write a fibonacci sequence in go in /tmp"
# Benchmark omp
time omp -p "Write a fibonacci sequence in go in /tmp"
# Check oMLX logs for cache and SpecPrefill stats
grep -E "SpecPrefill|cached" ~/.omlx/logs/server.log | tail -10Expected results on M4 Max 64GB with Qwen3.6-35B-A3B-MLX-8bit:
| Agent | Time | Notes |
|---|---|---|
| pi | ~16s | Minimal system prompt; tool calls dominate |
| omp | ~19–25s | SpecPrefill on large system prompt adds ~10s |
- pi docs: preferred agent; smaller system prompt, faster on local models
- pi providers
- pi custom models
- pi-mcp-adapter: bridges MCP servers to pi's tool interface
- omp / oh-my-pi: same codebase as pi; heavier default system prompt
- oMLX: local inference server for Apple Silicon
- oMLX agentic coding guide: hardware tuning, model profiles, agent config
- rtk: CLI proxy that filters/summarizes command output; wired into pi via
extensions/rtk-rewrite.ts