Configure the Harness Before Building One

One evening in September I put Hermes, an open-source agent runtime, in Docker on my Mac, connected it to Claude, and wired it to Telegram. I had my own AI bot on my phone. It worked. It also cost about $10 in API calls on Sonnet, in one evening. My Claude Pro plan costs $20 for the whole month.

The Insight

Hermes is good software. Nous Research built it as a runtime: one gateway for Telegram, Discord, Slack and more, a scheduler, memory, and a free choice of models. The catch is that a runtime has to be run. Owning it made me the platform engineer.

The setup record reads like a maintenance log. Hermes could only write files through workarounds: 118 terminal calls in one session. Six tools went out with every request, with nothing behind them. Fifty-five skills were loaded, two of them mine. I learned to type /new to reset the conversation, because every message carried the whole history and the bill grew with it.

I can’t prove which of those drove the $10. I think the weight did. On a metered API, every unused tool and every stale message gets paid for again on every call.

The Fix

I rebuilt my newsletter studio inside Claude Code, on the plan I already pay for. Claude Code comes with Pro and draws from the same usage limits.

Here is what I built. An operating manual that loads at the start of every session. Nine skills, one per step, from intake to retro. Two guards that run before every command and file edit, and block deletes, overwrites, and any attempt to approve its own publishing. Four scripts for the mechanical work, each with tests. Three gates where nothing moves without my yes. Plain-text files that say where every newsletter stands.

I thought I had engineered my own harness. I hadn’t. By Anthropic’s own description, Claude Code is the harness: it runs the loop, the tools, and the hooks that fire before every command. I configured one. That difference matters, because configuring is cheaper than building, and much cheaper than operating. Hermes gave me a runtime to look after. Claude Code gave me a harness to set up.

The Hard Truth

Claude does not do everything. Its scheduled tasks run in the cloud while the Mac sleeps, but they can’t reach a folder on my computer, so local jobs still need the Mac switched on. Hermes still wins for a Telegram bot, a webhook listener, or switching between models.

So before adding any layer, I ask one question. What does the tool I already pay for do, and what is the one limit this new layer removes? If I can’t name the limit, the layer hasn’t earned its place.

Sources

• Claude Pro at $20 a month, with Claude Code included and sharing the same usage limits (September 2026) — Claude
• Hermes: open source, one gateway for Telegram, Discord and Slack, a scheduler, any model — Nous Research
• Claude scheduled tasks run in the cloud but can't be tied to a folder on the computer — Claude Help Center
• Claude Code as the harness that runs hooks before every tool call — Claude blog