Goal: Find the exact task-complexity ceiling where a local Qwen3-MoE 30.5B running in Ollama, driven by the claw-code Rust harness, sustains autonomous multi-step tool use — and where it collapses.

This is not a vibe check. It is a reproducible, publishable boundary test: local sovereign compute vs. cloud API agents, in 2026, on real agentic workloads.

Motivation: If a cloud AI blackout occurs — data center loss, geopolitical disruption, infrastructure attack — you need a local fallback that actually works. This study tells you which model to trust and why.

Verdict: use open-claw (Qwen3-MoE 30.5B) as your primary local fallback.

• The 5.3% score gap is within noise. What is not noise: gemma4's P95 latency is 36 seconds. Under incident conditions, a model that freezes for 36s on a planning task is operationally useless.
• Multi-step planning FAIL on gemma4 is a hard blocker. Agentic loops require chained reasoning. A model that cannot plan 3 steps reliably will collapse on any real workload.
• open-claw's tool call emission failure is inconsistent across runs — likely a prompt sensitivity issue, not a capability gap. gemma4's planning failure is deterministic.
• open-claw is faster at median (3938ms vs 6047ms) despite being 4x larger. MoE routing wins.

On solar/salt-tank backup, open-claw draws ~4x more power. That is a real cost. But a fast, wrong answer costs more than a correct answer at higher wattage. If you have the power budget: run open-claw. If you are severely power-constrained and your workload is simple Q&A (not agentic), gemma4 suffices.

claw.exe  -->  ANTHROPIC_BASE_URL=http://localhost:8787
                    |
              ollama_proxy.py
              (Anthropic <-> OpenAI format translation)
                    |
              Ollama  -->  open-claw (Qwen3-MoE 30.5B)
                    running on YOUR machine. No cloud. No data center.

# 1. Build the harness
cd rust && cargo build --release && cd ..

# 2. Start the proxy (separate terminal)
python ollama_proxy.py --port 8787

# 3. Run the comparison
python compare.py --proxy http://localhost:8787 --save my_results.md

# 4. Interactive session with open-claw via claw harness
set ANTHROPIC_BASE_URL=http://localhost:8787
set ANTHROPIC_API_KEY=ollama
rust\target\release\claw --model open-claw

Requirements: Rust 1.70+, Python 3.11+, Ollama with open-claw and gemma4 pulled.

rust/                   claw-code Rust harness (build here)
  crates/
    api/                Anthropic API client + SSE streaming
    runtime/            ConversationRuntime, config, permissions, MCP, session
    tools/              Bash, ReadFile, WriteFile, EditFile, Grep, Glob, Agent...
    rusty-claude-cli/   REPL + one-shot prompt binary
ollama_proxy.py         Anthropic <-> Ollama format bridge
compare.py              Head-to-head evaluation suite
comparison_report.md    Raw results from April 2026 run
src/                    Python porting workspace (reference only)
tests/                  Python verification surface

This repository is not affiliated with or endorsed by Anthropic.