Source: raw/Hermes_+_Claude_Code_Is_The_Ultimate_Cheat_Code.md (auto-caption transcript)

Creator: Sharbel (YouTube @sharbelxyz) · URL: https://www.youtube.com/watch?v=hZqC6-pHPng · Duration: 14:11 · Uploaded: 2026-09-15 · No sponsor named; template offered free in the description.

Sharbel’s rule for running two agents together: Claude Code is the builder that sits with you and makes things, and Hermes is the operator that runs them on a schedule without you. A shared memory layer lets each know what you told the other. He opens with an unverified claim that the setup closed a $6,000 client overnight.

Key Takeaways

  • Opposite jobs. Claude Code works in bursts while you drive it and forgets the project when closed. Hermes lives on a server or spare machine, runs on schedules and is reached from your phone, but is not as strong a coder. “You build it once with Claude Code and you hand it to Hermes to run forever.”
  • Codex works as the builder too. If you already pay for ChatGPT, Codex plus Hermes covers the whole stack with no new subscription.
  • Three examples from his business:
    • His website: Claude Code built it; Hermes watches sign-ups and email open rates and, when a number slips, tells him and Claude Code what to change.
    • His videos: Claude Code built a video-editing skill. Hermes watches second-by-second retention and his delivery, writes him a brief, and feeds the same brief back so the editing skill improves.
    • Client services: Claude Code builds the websites and systems sold; Hermes keeps a pulse on every client.
  • Shared memory makes it one system. By default the two share nothing. Point both at the same memory account and user ID: in Claude Code through Connectors or two terminal commands, in Hermes by asking it to connect and giving it the API key. It adds to each tool’s own memory rather than replacing it. His demo: he told Claude his two favourite foods and Hermes answered correctly.
  • Memory options he lists: “Memo” (his pick, fast and free to start), Zep (stores what was true and when), Letta (the agent edits its memory in the open, so you can audit every node), Supermemory (which he calls simple and fully local). ^[ambiguous: the operator-use-cases video from the same creator calls Supermemory cloud-based]
  • Self-improvement. When something fails, Hermes works out why and rewrites its own instructions, so a correction becomes a lasting skill.

His five rules

  1. Build with the expensive model, run with the cheap one (or a local one).
  2. Tell Claude Code to build so Hermes can run it, from the start.
  3. Add “verify the result before you call it done” to every job.
  4. Gate anything scary: Hermes may do anything except send emails, spend money or delete without asking.
  5. Let Hermes call Claude Code itself on a schedule when new code is needed. If you use only Hermes, create a builder profile on your smartest model for Hermes to call.

Try It

  1. Pick one thing you already built with Claude Code and write a one-paragraph operating brief for Hermes: what to watch, what threshold triggers an alert, where to send it.
  2. Connect both tools to the same memory provider under one ID, then test it with a fact neither could guess.
  3. Add the verify line and your three gates to the Hermes job.

Open Questions

  • “Memo” is a caption spelling, possibly Mem0 ^[inferred]; not confirmed.
  • How Hermes invokes Claude Code (CLI, Codex app-server, MCP) is not shown.