The honest map

A local agent works very well for bounded tasks: explaining code, generating tests, creating scripts, refactoring a small component, or preparing a change plan. For large migrations, huge repositories, or long reasoning, the norm is to use a hybrid approach: local for privacy and drafts, cloud for difficult decisions or final review.

The practical foundation is Ollama. Its official documentation lets you run models on macOS, Windows, and Linux, and expose a local API for integrations. On Apple Silicon, MLX can be useful for experiments and performance with unified memory, but it doesn't turn any laptop into an unlimited workstation.

TerminalCode

# 1. Comprueba Ollama
ollama --version
ollama list

# 2. Descarga un modelo de código o propósito general
ollama pull qwen2.5-coder:7b

# 3. Prueba una tarea pequeña
ollama run qwen2.5-coder:7b "Explica este bug: un array vacío rompe un reduce sin valor inicial"

# 4. API local para integraciones
curl http://localhost:11434/api/generate -d '{
  "model": "qwen2.5-coder:7b",
  "prompt": "Escribe una función TypeScript que valide emails sin dependencias.",
  "stream": false
}'

Checklist before letting it touch files

Work on a branch: never test a local agent on `main`.
Define scope: which folder it can read, which files it can edit, and which commands it can execute.
Require a plan: before editing, have it list files, risks, and success criteria.
Verify cold: install dependencies and run tests after the change.
Log metrics: time, tokens, errors, and commands executed.

TerminalCode

# Flujo mínimo para trabajar con seguridad
git checkout -b agent-local/tarea-pequena

# Pide primero un plan, no cambios
# Después de aplicar cambios:
npm install
npm run lint
npm run build
git diff --stat
git diff

Official sources

Ollama Quickstart
Ollama: context length
MLX documentation