Skip to content
aiagentsetups.com

Setups · Featured

Auxten Wang's one-man company on Hermes

Auxten Wang (@auxten) · Catalog date October 17, 2026

Auxten Wang runs Hermes Agent 24/7 on a Mac mini and operates it from Discord. Scheduled jobs tune ads weekly, run a read-only anomaly check daily, and refresh App Store keywords and metadata every 14 days. Hermes also supervises Claude Code and Codex sessions in tmux. The same slides promote HandsOn, which they call open-sourced, and link his ClickMem repository.

Evidence. Sourced write-up

Goal
Run a one-man company's app operations, and supervise coding sessions, from Discord
Use
Coordination
Platforms
Hermes, Claude Code, Codex, Cursor
Crew size
Multiple agents

Roles

The slides describe one Hermes agent on the Mac mini, a pile of Claude Code sessions, and a Codex pane beside them. They do not give a headcount. A personal Codex subscription is also the backend model for Hermes. The slides do not say that subscription and the Codex pane are the same process. Cursor is named for daily dev and for the laptop side of a handoff. The slides do not say Hermes supervises Cursor. cc-supervisor is a dashboard Hermes says it started, not a separate named agent in this entry.

  • Hermes Agent

    Owns. The 24/7 agent on the Mac mini. It takes Discord messages, runs the scheduled jobs, reports back to Discord, and supervises coding sessions in tmux.

    Constraints specific to this role

    • In the foreman reply on the slides, Hermes says it will not press Enter to commit unless explicitly asked. The slides do not apply that sentence to every routine, including the App Store refresh that ships a new version.
    • The daily anomaly inspection is read-only.

  • Claude Code

    Owns. Coding sessions Hermes supervises. The foreman diagram shows Claude Code in tmux pane 0:0. The stack slide also lists Claude Code for daily dev. The slides do not say how many sessions run.

  • Codex

    Owns. The foreman diagram shows Codex in tmux pane 0:1. Separately, the stack slide says a personal Codex subscription is Hermes's backend model, with MiniMax M3 via OpenRouter as backup.

Routines

  • Weekly ad-campaign optimization

    Sun 08:00-09:30. No timezone is stated.

    Per-app keyword and bid tuning across his apps, then one roll-up report to Discord.

  • Cross-app anomaly inspection

    Daily 08:00. No timezone is stated.

    Read-only check of installs, trials, revenue, and LLM errors. The slides say it pings him only if something is off.

  • Emby pick of the day

    Daily 10:00. Another slide says every day at 10am. No timezone is stated.

    One title from his Emby server, with a reason, and not something he has already seen.

  • App Store refresh

    Every 14 days

    Refresh App Store keywords and metadata, then ship a new version. The slides do not say what ship includes.

  • Anime download

    Sat 15:00. No timezone is stated.

    Download the week's new anime episodes to his NAS.

Guardrails

  • Hermes runs as its own macOS user on the Mac mini. The slides call that blast-radius containment and also say to use Time Machine.
  • The slides say Hermes acknowledges a task before any model call, and that every message gets a final result, success or failure.
  • ClickMem is explicit-only. An agent commits a refined conclusion with clickmem_remember, or he imports curated docs such as AGENTS.md, CLAUDE.md, or rules. Raw chat is not auto-injected. A memory that contradicts an older one is flagged for him to resolve.
  • The daily anomaly inspection is read-only and pings only when something is off.

Handoffs

Auxten sends tasks in Discord, and Hermes reports cron results there. For coding, the slides say Hermes opens tmux and supervises Claude Code and Codex panes: it summarizes output, chases progress, and quality-reviews. Hermes says cc-supervisor refreshes every pane every 30 seconds. Work built on the laptop in Cursor, Claude Code, and Codex is handed to Hermes on the mini. The slides do not describe a file format for that handoff. They say context used to evaporate, which is the problem ClickMem is presented as solving.

Recorded handoff
  • Hermes Agent
  • Claude Code
  • Codex

Human operator

Auxten sends tasks in Discord, and Hermes reports cron results there. For coding, the slides say Hermes opens tmux and supervises Claude Code and Codex panes: it summarizes output, chases progress, and quality-reviews. Hermes says cc-supervisor refreshes every pane every 30 seconds. Work built on the laptop in Cursor, Claude Code, and Codex is handed to Hermes on the mini. The slides do not describe a file format for that handoff. They say context used to evaporate, which is the problem ClickMem is presented as solving.

Results

What worked
The slides show Hermes replying that a May 14 message synced 27 custom skills from ~/.openclaw/skills/ to ~/.hermes/skills/openclaw-imports/, with no mismatch. They say the agent found a bug in the platform it runs on, wrote a one-line fix and two tests, opened the pull request under its own GitHub account, and three human reviewers approved it. They also show the scheduled Discord reports described above.
What broke
The slides say context evaporated when work moved from Cursor, Claude Code, and Codex on the laptop to Hermes on the Mac mini, and that auto-capture made memory worse: a one-week vegetarian trip became a permanent diet, and a dropped Firebase idea kept appearing in designs. They say Hermes could not click macOS security dialogs after it pushed an update, because macOS blocks software from clicking those dialogs. They also say a plugin reload caused a restart loop and gateway death, which is the bug the agent then patched.
Lessons
The closing slide lists five points: infrastructure over prompts; one sentence can move skills and memory; bad memory is worse than no memory; an instant acknowledgement and a closed loop earn trust; treat the agent like a coworker, and containment is the safety.

Prompts

  • Migrate skills

    Learn all the skills I defined over on OpenClaw.

    Shown as a Discord message dated May 14. The slides do not give the year. They do not include the skill files.

  • Foreman

    Be a foreman: open a tmux and manage my pile of Claude Code sessions — summarize key output, chase progress, and quality-review (catch laziness / hallucinations / fake tests).

    Shown as the message that started cc-supervisor. The diagram also labels a Codex pane. The slides do not include the cc-supervisor skill file.

  • App status check

    Check the online status of my apps.

    Shown as a Discord message sent at 08:00, with a reply at 08:17. The slides do not say whether this message is the daily anomaly cron.

Source

Notes

The catalog date is the day this entry was drafted. The slides are labeled PyCon Singapore 2026, The Agent Track. They do not give a calendar date. The skill-migration message is dated May 14 and does not give a year.

The slides also promote his own projects. The HandsOn slide says it is open-sourced, one MCP interface with macOS, Rock 5B, PiKVM, and NanoKVM backends, and the closing slide links github.com/auxten/handson. The same closing slide links github.com/auxten/clickmem and github.com/chdb-io/chdb. He is identified as the creator of chDB. The slides do not state a license for ClickMem or chDB. Hermes Agent is attributed to Nous Research, MIT, shipped February 2026. Those slides also say it has about 195k GitHub stars and that Hermes leads by daily token volume. Those are claims on the slides about the product.

His stack slide says: Discord on phone and desktop; Hermes Agent on a Mac mini, 24/7; a personal Codex subscription at about $200 a month, with MiniMax M3 via OpenRouter as backup; ClickMem on chDB and Qwen3-Embedding-0.6B; HandsOn, a Rock 5B IP-KVM he prices at $182, with USB HID, HDMI capture, and mutual passwordless SSH so that if one machine dies the other revives it. Total draw is given as 7 watts. The ClickMem slides say recall is similarity plus a filter, scoped by project, privacy, and source, with chDB local and ClickHouse at scale, and they say there is no server-side LLM. Recurring cost is given as about $200 a month for the Codex subscription he already pays. The product overview says Hermes can run in Docker, SSH, or serverless backends. His stack slide does not say he uses those backends.

The one-rule slide says to give an agent its own email, user account, GitHub, and machine. He says his runs as a separate macOS user. He writes that if it does something destructive, and that it will, the separate user is the sandbox. The slides do not describe a specific destructive incident. The pull-request slide says the agent opened the pull request under its own GitHub account. The slides do not say he gave it a separate email.

A sample ASA report on the slides is labeled "Daily ASA report" and captioned "Weekly ad optimization, per app," timestamped Sun 09:30. The figures printed there are: yesterday, spend $106.63, 32 installs, CPI $3.33; last 7 days, spend $440.16, 104 installs, CPI $4.23; since re-enabling, installs 2, 27, and 32 on the dates printed as 6/14, 6/15, and 6/16. A daily-watch sample at 08:17 says nothing critical, three medium anomalies, and gives one app's GA4 engagedSessions as 10 against a 7-day average of about 35.8, down about 72 percent, which the slide calls a likely tracking artifact. These are figures shown on the slides, not measurements by this catalog. No timezone is stated. The app names are not listed. The status-check example says Hermes looked at Langfuse, GA4, RevenueCat, and ASA.

The Emby examples name shows and record taste feedback, including marking Billions as watched and stopping recommendations of Day of the Jackal. The Saturday anime job does not name the shows. The slides do not include the cron definitions or the ad-ops skill files.