Uses

Tools and stack I actually use day to day. This is a practical snapshot, not a forever list.

last updated: 2026-05-11

// AI + model stack

Coding: OpenAI Codex 5.5 as the main heavy-lift coding model
Planning/review: Claude Sonnet/Opus when I want a second opinion or sharper critique
Fast/cheap tasks: Gemini + Haiku/Sonnet for lower-stakes routing, summaries, and quick checks
Voice/transcription: OpenAI + ElevenLabs, with Whisper still useful when I want local transcription
Image generation: GPT Image 2
Provider routing: Hermes routing across OpenAI, Claude, Gemini, OpenRouter, and local-ish tools where useful

Harness: Hermes Agent. OpenClaw did the first messy tour of duty; Hermes is the current system.
Assistant shape: Giles as the main operator, with thinner specialist profiles only when they earn their keep. No swarm theatre for the sake of it.
Memory: Hermes memories for durable facts, session search for recall, and markdown ledgers for curated context that should survive model churn.
Skills: Reusable Hermes skills for coding, research, Gmail, UK public-data CLIs, tracking, meal planning, writing, and project workflows.
Crons/watchdogs: Deliberate scheduled agents: briefings, backups, email sweeps, proof-of-life checks, and change-driven alerts. No pointless heartbeat spam.
Channels: Discord-first for day-to-day ops, with Telegram and other surfaces used when the workflow fits.

Hermes Console: Read-only local UI for Hermes runtime health, sessions, cron, skills, memory, files, config, logs, and usage. Built because digging through five places gets old fast.
llm-usage: Small CLI for tracking token usage across Claude Code, Codex, and OpenRouter. Boring visibility, extremely useful.
OpenClaw Blackbox: Legacy run-inspection work from the OpenClaw era: failed/stalled/expensive/weird agent runs turned into deterministic reports.
Local CLIs: gog for Google Workspace, bird for read-only X/Twitter, plus UK-data CLIs for TfL, rail, Companies House, Parliament, and similar agent-friendly lookups.
Agent email: Gmail/Google Workspace through local authenticated tooling; AgentMail when a dedicated agent inbox is the right abstraction.

Email chief-of-staff: Inbox sweep -> labels, attention list, safe triage, and drafts when needed. Destructive/outbound actions still need approval.
Morning/weekly briefings: Calendar, tasks, notes, memory, and project context -> actionable focus list
Research reviews: URLs, repos, papers, videos, and X bookmarks -> receipts, notes, and next actions
Project ops: Plans, coding-agent delegation, PRs, release checks, changelog hygiene, and post-merge verification
Personal tracking: Discord natural-language logs -> structured tracking data for fitness, habits, and experiments
Dump loop: Raw capture -> curated daily ledger -> reusable memory/skills when the signal is worth keeping

Editor/IDE: VS Code, with terminal-native agents doing a lot of the heavy lifting now
Terminal: Ghostty on macOS, Kitty on Linux
Shell: zsh + powerlevel10k
Notes: Markdown files in git + wiki for durable knowledge; Notesnook for personal notes
Browser: Brave + Chrome

Messaging: Agent-first across Discord and WhatsApp/Telegram-style surfaces, with channels treated as workflows rather than inboxes to babysit
Calendar: Google Calendar, usually handled through the agent
Tasks: Google Tasks and markdown plans, mostly operated through agents rather than manually fiddled with
Focus: pomo CLI pomodoro app