Date: April 15, 2026
For: Guillaume Descoteaux-Isabelle (jgwill)
Research Method: 3-cycle multi-agent orchestration (10 claude-opus-4.6 sub-agents)
| # | Document | Description |
|---|---|---|
| 1 | OpenClaw Plugins for Local AI Workflows | OpenClaw plugin ecosystem — which plugins to install for local AI workflows (Ollama, HuggingFace, Copilot, Google, Perplexity), multi-provider routing, Hermes compatibility |
| 2 | Mac Mini for Local AI Inference — Hardware Scenarios | Mac Mini hardware for inference — Scenario A (minimal, ~$999) vs Scenario B (evolved, ~$2,399+), model memory requirements, benchmarks, pricing |
| 3 | Mac Mini for Local AI Training — Hardware & Workflow Guide | Mac Mini hardware for training — MLX/MLX-LM frameworks, QMD model fine-tuning, LoRA persona adapters, data sovereignty protocols, automated weekend pipelines |
| 4 | GitHub Copilot + Google Plugin Capabilities | GitHub Copilot + Google Plugin deep dive — model catalog, premium request quotas, Google multi-capability contracts, cost analysis, cloud+local routing |
| Document | Description |
|---|---|
| AGENTS.md | Full orchestration report — what agents ran, what they found, corrections applied, lessons learned |
| Agent | File | Focus |
|---|---|---|
| A | 00-initial-research/agent-a-plugins.md | OpenClaw plugin ecosystem |
| B | 00-initial-research/agent-b-mac-inference.md | Mac Mini M4 inference specs |
| C | 00-initial-research/agent-c-mac-training.md | Apple Silicon training capabilities |
| D | 00-initial-research/agent-d-qmd-models.md | QMD HuggingFace models analysis |
| E | 00-initial-research/agent-e-copilot-google.md | Copilot + Google plugin |
| Reviewer | File | Covered |
|---|---|---|
| Track 1 | 01-review/review-track1-plugins-inference-copilot.md | Agents A, B, E |
| Track 2 | 01-review/review-track2-training-qmd.md | Agents C, D |
| Reviser | Output | Corrections Applied |
|---|---|---|
| 1 | RESULT-01 + RESULT-04 | 15 items from Track 1 review |
| 2 | RESULT-02 | Pricing, specs, benchmarks from Track 1 |
| 3 | RESULT-03 | 2 BLOCKING issues + 12 items from Track 2 |
What Mac Mini should I buy for inference?
→ See RESULT-02. Minimal: M4 24GB ($999). Large models: M4 Pro 64GB (~$2,399).
What Mac Mini should I buy for training?
→ See RESULT-03. Sweet spot: M4 Pro 48GB (~$1,799). Max capability: Mac Studio M4 Max 128GB (~$3,699).
Which OpenClaw plugins do I need?
→ See RESULT-01. Essential: ollama-provider. Useful: copilot-provider, google-plugin. Skip: minimax, moonshot.
Can I fine-tune QMD's models locally?
→ See RESULT-03. Yes, with caveats. Query expansion model has existing pipeline (CUDA-dependent). Embedding model fine-tuning is feasible but GGUF conversion is unverified.
What does the Google plugin do?
→ See RESULT-04. Multi-capability: Gemini LLM, image gen, video gen, web search grounding. Requires separate API key.