Hermes as orchestrator with Claude CLI as reviewer and Codex CLI as builder - using existing subscriptions at zero additional API cost. The subscription routing insight: profile-based Claude spends API balance; Claude CLI uses the subscription directly. Live test result was better than Ralph Loop.

Watch on YouTube

The Architecture

Hermes (orchestrator + memory + loop)
  -> assigns Codex CLI (builder)
Codex builds code
  -> Claude CLI reviews (claude -p flag)
  -> Returns PASS or FAIL with notes
If FAIL: Codex picks up review notes and fixes
If PASS: Hermes does final verification against spec
Done: Hermes reports completion

The Subscription Routing Insight

Critical: when you add Claude as a Hermes profile (using Hermes API routing), it pulls your extra API balance - NOT your Claude subscription credits.

Running Claude via the claude CLI uses the subscription directly, at no additional API cost beyond the subscription.

Same applies to Codex: Codex CLI uses your ChatGPT/Codex subscription. Calling Codex via API from Hermes spends API tokens.

Rule: use CLI where possible to route work through subscriptions. Add as Hermes profiles only when you need Hermes native features (memory, skill routing) for that agent.

Setup

Verify both CLIs are installed:

claude --version
codex --version

If not installed, tell Hermes: "Install Claude Code CLI and Codex CLI so I can log in with my subscription credentials."

Builder and Reviewer Soul Files

Create two specialized profiles: builder and reviewer.

builder soul.md: "You are the builder. Use the Codex CLI exclusively. Do not use Hermes native tools or models. When assigned a task from the Kanban board, run the Codex CLI with the task spec. Return the result to the orchestrator."

reviewer soul.md: "You are the reviewer. Use the Claude CLI exclusively (claude -p). Review code against the spec and agents.md. Return PASS or FAIL with specific, actionable notes. Do not use Hermes native tools."

These profiles act only as CLI triggers. They do not use Hermes inference.

Why Not Just Hermes Profiles?

Four reasons to use CLI workers instead of Hermes profiles:

1. Subscriptions: CLI uses subscription credits; profile uses API balance
2. Separation of concerns: builder builds, reviewer reviews - no ambiguity
3. Native ecosystems: Codex brings its MCPs and plugins; Claude brings its own
4. Hermes retains its strength: memory, context, loop management, scheduling

Live Test Result

Same Kanban app built in the /goal comparison video, now with Hermes + Codex builder + Claude reviewer:

• Drag and drop: worked (Codex 5.5 + Claude Sonnet review)
• Cards: create, edit, delete, persist on refresh - all functional
• Claude review step caught issues and Codex fixed them before final report
• Better than the Ralph Loop test - Nemanja was surprised

Final report: "Spec completed. Build completed. Review fixed. Codex builder (5.5). Claude reviewer (Sonnet)."

Beyond Code

This pattern is not limited to code. Any workflow you run in Claude Code or Codex can be orchestrated by Hermes:

• Hermes initiates a content pipeline; Claude handles scripting; Codex handles data processing
• Hermes initiates an analysis workflow; Claude handles reasoning; Codex handles batch file processing
• Whatever automation you have already built in Claude or Codex - Hermes can operate it and run the loop

Hermes excels at: operating existing CLIs, maintaining loop state, providing memory across sessions. It does not need to build or reason at the level of Claude/Codex to be the most valuable piece of the system.

Fleet Mapping

Hermes Component	Carlos Fleet Equivalent
Hermes orchestrator	Carlos (loop management, sequencing, memory)
builder profile (Codex CLI)	Merlin (builds)
reviewer profile (Claude CLI)	Dan or reviewing lead
Hermes Kanban	Carlos task tracking
Subscription routing insight	Claude Max quota preservation rule