Run Pi locally on Mac.

Step 1 · Launch via Ollama

One command to start.

$ ollama launch pi

Step 2 · Pre-flight memory check

Will your Mac fit it?

Pi pairs best with Qwen 2.5 Coder 32B at Q4_K_M (~19.0 GB). Add ~4–8 GB on top for the agent's working set, plus headroom for the rest of your dev stack. Recommended: 24 GB Mac (minimum 16 GB).

$ devpulse ai --before-load 19456 --auto-clean
before: Won't fit — 6.4 GB short
  - unloaded idle ollama model: qwen2.5:7b (4.2 GB)
  - killed 4 zombie procs (612 MB reclaimed)
after:  Fits comfortably — 3.1 GB headroom

# safe to launch
$ ollama launch pi

Why this matters: Lighter footprint — runs comfortably on most Macs paired with a small-to-mid model.

Step 3 · Babysit long sessions

Don't lose progress to OOM mid-task.

Agent runs that span hours hit memory pressure as context grows. devpulse babysit watches and auto-cleans without crashing the session.

# in another tmux pane
$ devpulse babysit --target-free-mb 8192 --json > pi.log &

# emits NDJSON tick / cleanup events you can checkpoint on:
{"event":"tick","tickNum":47,"availableForAIMB":7200,"pressure":"free<8192MB",...}
{"event":"cleanup","reasons":"free<8192MB","reclaimedMB":5400,...}

Tips for Pi

What we've learned running it.

• Lightest agent footprint here — runs comfortably on 16 GB Macs paired with an 8B-14B model
• Add capabilities deliberately via extensions (web search via @ollama/pi-web-search, etc.) — keeps RAM usage proportional to what you actually use
• When swapping providers in `~/.pi/agent/models.json`, use `devpulse status` to confirm the new model fits before invoking
• Plays well with `devpulse watch --json` since Pi's predictable footprint makes pressure events easy to attribute to other parts of your stack

Recommended models

Best local pairings for Pi.

Related agents