ai-orchestrator cuts Claude spend with Ollama

The idea is practical, not flashy: most coding work is orchestration, formatting, and repair, so a tiered model stack can save real money if the handoff boundaries are clean.

• –The strongest part is task decomposition: Claude writes the spec, local models implement against concrete files, and deterministic checks close the loop
• –This works best when the spec is precise; local coder models still tend to stumble on ambiguous architecture or broad refactors
• –The claimed savings are believable for repetitive TS/Python work, but the win depends heavily on project structure, test coverage, and how often auto-fix loops churn
• –Bash, `jq`, and `curl` keep the system lightweight and portable, which makes it easier to adopt than a heavier agent framework
• –The broader pattern is useful: cloud for judgment, local for throughput, especially when token costs matter more than absolute elegance