Claude Code on a laptop-local model

A private, offline coding agent that keeps Claude Code's workflow.

• Claude Code's 28-tool prompt (about 35k tokens) overflowed the local model's 32k context.
• The model's chat template crashed on multi-block system prompts.
• The standard benchmark evaluator scored every task 0 on macOS.

1. A context-budget proxy that keeps 5 tools with full schemas and swaps in a compact system prompt.
2. Patched chat template and cache-safe headers, so repeat turns reuse the prefix cache.
3. Benchmarked a mixture-of-experts model against a dense one: quality tied within 2.5 points, the MoE decoded about 7x faster, so it became the daily driver.
4. Replaced the broken evaluator with a subprocess scorer, cross-checked by an independent harness.
5. SHA-256-verified model downloads.

• 83% fewer (open receipt)Receipt Verified 2026-10-09Cut Claude Code's per-request prompt from about 35.4k to about 6.1k tokens with a 5-tool context-budget proxy.Period2026-10-09MethodSame captured 28-tool request before and after; tokens estimated as characters / 4.SourceSession capture prompt tokens per request.
• HumanEval+ 87.2% (open receipt)Receipt Verified 2026-10-09The chosen open-weights model scored 87.2% on HumanEval+ (143/164) and 70.4% on MBPP+ (266/378).Period2026-10-09MethodGreedy decoding; scored by my subprocess scorer, cross-checked by an independent harness.SourceBenchmark results on the chosen model; MBPP+ 70.4%.
• 70.4 tok/s (open receipt)Receipt Verified 2026-10-09Decoded at 70.4 tokens per second on an Apple M4 Max.Period2026-10-09Methodllama-bench tg128, Q6_K, flash attention.SourceBenchmark output decode and 2.7 s (open receipt)Receipt Verified 2026-10-09A Claude Code tool call through the proxy took 2.7 seconds (33.4 s on the dense fallback).Period2026-10-09MethodTimed live probe.SourceSession capture per tool call on an M4 Max.