Local coding-agent setup for Apple Silicon macOS:
local-agent → vanilla Pi → Rapid-MLX → Qwen3.6 MLX model
git clone <repo>
cd local-agentic-dev
./install.shThe installer installs Rapid-MLX and Pi, downloads the model, configures Pi, installs the adaptive-thinking extension, and installs the local-agent command. It is safe to run again.
local-agentThis starts Rapid-MLX when needed, opens Pi with the local model already selected, and stops the server when Pi exits.
local-agent status
local-agent logs
local-agent helpThe canonical settings are in config/local-agent.conf:
model unsloth/Qwen3.6-35B-A3B-UD-MLX-4bit
endpoint http://127.0.0.1:8000/v1
context 98304
output 24576
Rapid-MLX uses the Qwen3.6-compatible qwen3_coder_xml tool parser and qwen3 reasoning parser. The default 98,304-token context and 24,576-token output are intended for large, complex enterprise web repositories; reduce them if unified memory is tight. Serving stays single-sequence with a small admission queue so a long coding session remains responsive. Requests may contain up to 32 MiB of tool/project data and can run for up to 30 minutes. Speculative decoding/MTP is disabled for this checkpoint.
All runtime values in config/local-agent.conf accept environment overrides, for example CONTEXT=32768 MAX_OUTPUT=8192 PI_THINKING_LEVEL=medium local-agent for a lower-memory run. PI_THINKING_LEVEL defaults to high and can be set to off, low, medium, high, or xhigh. Pi also loads the pi-adaptive-thinking extension, which lets the model temporarily raise or lower the Pi thinking level through a native tool; its global configuration is ~/.pi/agent/adaptive-thinking.json.
Pi configuration lives in ~/.pi/agent/. Models are stored in the Hugging Face cache, normally ~/.cache/huggingface/hub/. Logs and runtime state live in ~/.config/local-agentic-dev/.
Run the installer again:
./install.sh./uninstall.shThis removes project configuration and runtime state, restores Pi backups, and preserves downloaded models and installed packages.