Going hybrid: cloud + local

You now have two worlds at hand. On one side, the big vendors’ frontier models, taken on a subscription with their coding agent: Claude with Claude Code, GPT with Codex, Grok with Grok Build, Gemini with Antigravity CLI. On the other, models that run for free at home (Ollama). The wrong question is “which one to choose?” The right answer is: both, at the same time, each in its place.

That’s the hybrid setup, and it’s the most effective way to work with a mini-PC. The machine orchestrates: it runs the agents and keeps your projects and scripts. The frontier plan, at a fixed price every month ($20 for Claude Pro or ChatGPT Plus, $30 for SuperGrok, €21.99 for Google AI Pro, prices checked on October 1, 2026), does the bulk of the reasoning. The local model takes the confidential, the repetitive and the offline. Tiers, limits and what each vendor allows are compared in Frontier models on a subscription.

The examples in this guide put Claude Code in the conductor’s seat, because it’s the most battle-tested setup; it carries over as is to Codex or OpenCode. The cloud agent decides who does what, and delegates the bulk work to the local model.

Why mix rather than choose

The two worlds have opposite strengths, and that’s exactly what makes them complementary:

• The cloud (Claude, GPT, Grok, Gemini) is unbeatable on hard reasoning, long tasks, reliable tool use, architecture. But plans have caps (per 5-hour window and per week), the API is billed per token, and your data goes to a third party.
• Local (Ollama) is free to use, private, available offline, and largely good enough for bounded and repetitive tasks. But it tires on long agentics and cutting-edge reasoning (we talked about it bluntly in Choosing your model, and the Quelle IA local models ranking, in French, puts numbers on the gap).

The hybrid setup takes the best of each: you keep cutting-edge intelligence where it really matters, and you knock down cost and data leaks on everything else, that is, 80% of the volume.

Claude Code as orchestrator: how it works

The thing that makes this possible: Claude Code can run commands. And Ollama exposes a dead-simple local API on http://localhost:11434. So Claude Code can call your local model, via a curl, a script, or a small tool, exactly as it would call any other command.

Concretely, you tell Claude: “for this bulk task, don’t do it yourself, delegate it to the local model via the Ollama API.” It writes the script that loops over your files, hits the local model for each one, and brings back the consolidated result. It keeps the big picture; local does the grunt work.

# The basic move Claude Code orchestrates: call the local model (here gemma4:26b, fast)
curl -s http://localhost:11434/api/generate -d '{
  "model": "gemma4:26b",
  "prompt": "Summarize this file in 3 bullets: '"$(cat report.md)"'",
  "stream": false
}' | jq -r .response

Four concrete ways to route the work

Wiring both up, in practice

# Hybrid strategy
- For repetitive, bulk tasks (rewriting, classification,
  summaries, boilerplate generation), delegate to the local model via the
  Ollama API (http://localhost:11434), don't do them yourself.
- Keep complex reasoning, architecture, and review for yourself.
- Code and data marked "sensitive": local model only.

# a Claude Code session with a local brain
ollama launch claude --model qwen3.8:27b