Tooling Decisions
Coding Harnesses
- Codex (GPT-5) — more creative, finds bugs others miss (discovered the GHL private API). Use for reverse-engineering and greenfield spikes.
- Claude Code (Opus 4.6) — more consistent, better instruction-following, superior for multi-file refactors and sustained sessions. Default choice.
- Migration away from OpenClaw forced after Anthropic removed third-party OAuth for harnesses.
Local Models
- Gemma 4 (26B) — runs on 3090 Ti, 24GB VRAM, ~40 tok/s, intelligence score ~39. Good for ICP filtering, batch classification, anything privacy-sensitive.
- Cloud fallback: Gemini 2.5 Flash for cheap/fast, Claude Sonnet 4.6 for quality.
Browser Automation
- Pinch Tab — ~5x token savings vs Playwright MCP for snapshot-heavy workflows. Use for repetitive UI tasks.
- Playwright MCP — when you need full scripting, storage state export, or network interception.
- Chrome DevTools MCP — pixel/form/API reverse engineering on live sites.
MCPs in Active Use
- Stripe MCP — automated invoicing, usage-based billing
- Plaid MCP — bank + CC feed for Claude-driven finance decisions
- Chrome DevTools MCP — API discovery
- Context7 — library docs
Password Management
- Vault Warden — self-hosted Bitwarden-compatible. Chrome extension works unchanged. One less SaaS bill, zero trust in a third party for credentials.
Visualization
- Mermaid — inline in markdown, LLM-friendly
- Excalidraw — whiteboard sketches, LLM can export/import
- Obsidian — personal knowledge base with Mermaid + dataview
- NOT Miro — not LLM-friendly, can't be version-controlled
The Rule
Prefer tools LLMs can read, write, and diff. Visual-only tools (Miro, Figma-only flows) become black boxes.