DevPulse → Local AI → Agents

Coding agents · running locally

Local coding agents you can run on Mac.

Local AI on a Mac splits into two tiers. Model runtimes (Ollama, llama.cpp, LM Studio, MLX) load weights and expose inference. Agent runtimes (the ones below) sit on top and orchestrate multi-step work — Claude Code, Codex, OpenCode, Copilot CLI, Aider, OpenClaw, and more. Each one solves the same problem: load a 32B–70B model, point it at your repo, watch it work.

And each one hits the same wall: OOM mid-task when Chrome, Docker, or a stale dev server eats the unified memory the model needed.

Below: the 9 agents we cover with their memory profile and DevPulse pre-flight commands. Click any agent for the full setup.

AgentRecommended modelMin RAMOOM risk
Claude Code →
Anthropic's coding agent with subagents, tool use, and multi-file edits.
llama3.3-70b64 GBHigh risk
Codex →
OpenAI's open-source coding agent for the terminal.
qwen2.5-coder-32b48 GBHigh risk
OpenCode →
Open-source coding agent with strong local-model support.
qwen2.5-coder-32b48 GBHigh risk
GitHub Copilot CLI →
GitHub's AI coding agent for the terminal.
qwen2.5-coder-32b32 GBMedium risk
OpenClaw →
Personal AI assistant with 100+ skills — runs locally via Ollama.
qwen3-32b32 GBMedium risk
Hermes Agent →
Self-improving open-source agent with persistent cross-session memory.
llama3.3-70b64 GBHigh risk
Droid →
Factory's terminal-first coding agent — top Terminal-Bench score, IDE integrations.
qwen2.5-coder-32b48 GBHigh risk
Pi →
Minimal, aggressively extensible coding agent CLI with native Ollama support.
qwen2.5-coder-32b24 GBLow risk
Aider →
AI pair programming in your terminal — works with local Ollama models.
qwen2.5-coder-32b48 GBHigh risk

The pattern

Same fix across every agent.

Whichever agent you pick, the operational problem is identical: your stack ate the unified memory before the model got a slice. DevPulse's pre-flight check + auto-clean is the universal fix.

Before launch

devpulse ai --before-load <MB> --auto-clean reclaims idle Ollama models, kills zombie LSPs, then re-evaluates. Exit 0 = safe to launch.

During the run

devpulse babysit --target-free-mb 8192 --json watches free memory + battery + swap. Auto-cleans when pressure builds, emits NDJSON checkpoint signals.

If something goes wrong

Fix Ollama OOM on Mac → covers the most common failure modes and the diagnostic commands to find them.

Stop letting your stack OOM your local agents.

DevPulse is free, native, and uses less RAM than this webpage.

Download for macOS

macOS 14+ · Apple Silicon & Intel · Free during launch