The 7 Changes That Make Hermes Agent Up To 95% Cheaper To Run
A free audit plugin took my overhead from 40,000 tokens to 19,200 before I changed a single model. Here's the full setup, step by step.
The full walkthrough is on YouTube, with the Token Optimizer dashboard on screen as I go through it. Watch it here: https://youtu.be/K7rooUY4-FQ
If you use Hermes Agent every day, a lot of what you pay for never touches the answer. You pay for what the model reads and what it writes, and most people only ever look at the writing side.
Every message you send arrives with the system prompt, the conversation so far, your memory files and the instructions for every tool. So a one-line prompt turns into thousands of tokens before Hermes has done anything at all.
My own first message, one short line at the start of a session, used 18,900 tokens. That's 7% of a 272K context window, with the system prompt taking 2.6% and the tool definitions taking 3.7%.
The changes below keep the quality where it matters, and a couple of them make the answers sharper.
By the end of this guide you'll know:
•
1. Token Optimizer: see what every message costs
Token Optimizer is a free GitHub plugin, and I run it on both Hermes and Claude Code. You don't need an install guide for it. Paste the link into Hermes, ask it to run it, and it hands you a full report.
Can you run this? https://github.com/alexgreensh/token-optimizer
The report shows your overhead, your quick wins and which tools eat the most tokens. On mine the top two were delegation and connections, and I didn't even know what they were.
It also flagged "many unused skills", which is change 2.
The Manage tab has a handful of settings, and two of them go straight at wasted tokens. Loop detection catches the agent when it gets stuck retrying the same thing, which burns tokens fast.
Delta mode stops the agent rereading a whole file after it edits it, so the agent changes only the part that was wrong.
If you want a faster check before the full audit, send Hermes a quick hello and then type /context. You'll see how much of the window is gone before you've asked for anything, and what's taking it.
2. Unused skills and MCPs: switch off what you never call
Hermes offers a long list of skills during setup, and it's easy to tick ones you will never use.
You don't need to delete every tool, though. Installed doesn't mean fully loaded on every message, so cut the ones you never call and leave the rest.
My setup shows 5 MCP servers and 184 skills, and I've already disabled 63 of them. Ask Hermes for the list, read through it, then tell it what to remove.
Can you show me a list of the MCPs and skills I've got installed here for Hermes?
Can you remove these skills and these MCPs: [names from the list]
3. The 40% rule: clear context before the model loses the middle
Skills load at the start of a session, but the conversation keeps growing after that. Every request you send includes the entire conversation in your current chat. So a long session costs more on every message, and the model starts losing information from the middle.
I learned this going through the Claude Code Certified Architect programme, and I passed the exam a couple of weeks ago. Don't wait until the window is 60, 70, 80 or 90% full. Clear it at 40 to 50% at the most.
At about 40%, I have the agent write a handoff file, then I clear the session and start the new one from that file.
That's different from compaction. Compaction compresses the session for you, and you get no say in what survives.
This session is at about 40% context. Before I clear it, write a file called handoff.md with:
- what I was working on and why
- the decisions made so far
- the files that changed
- the next step, written so a fresh session can start on it without asking me anything
Keep it under one page. I'll start the next session from this file.
Keeping Hermes online when your laptop is closed
A clean session only helps if the agent is there when you need it. Hermes works like a real agent when it stays online with the laptop shut. I run mine on a Hostinger VPS, so it can talk to me on Telegram from my phone.
It's a managed install with one-click setup and Telegram pairing built in, and the security is handled for you. It also takes weekly backups, and you can take a manual snapshot before you update Hermes or change the setup.
The 24-month plan gives the biggest discount, and the code GROWTHLAB10 takes another 10% off at checkout.
4. Ponytail: less code, fewer output tokens
The first three changes cut what Hermes reads. Ponytail cuts what it writes.
Ponytail makes your agent behave like the laziest senior developer in the room. Agents tend to write far more code than a job needs, and Ponytail pushes them toward the 20% of the code that does 80% of the work.
Given the same agent and the same task, it writes fewer lines, uses fewer output tokens, costs less and finishes faster. It was the number one repository at one point, and I run it in both Claude Code and Hermes.
Install it the same way as Token Optimizer, and it works in the background with nothing else to do. It also comes with commands to set how hard it pushes, review your code or audit a whole repo.
Can you install this? https://github.com/DietrichGebert/ponytail
5. Your agents file: point to other files instead of pasting them in
Back on the reading side, the next place to look is your agents file, which loads into every session. In Hermes that's AGENTS.md and in Claude Code it's CLAUDE.md, and both get bloated as you keep adding instructions.
So treat it as a map. Instead of pasting your voice rules or project notes into it, give the path to the file that holds them. The agent then opens that file only when the job needs it.
When I asked Hermes to review my agents file, it came back short, with a couple of stray lines to tighten. My CLAUDE.md is long on purpose. It loads in every Claude Code session I run, and it carries the routes to every folder in my workspace.
Can you review my agents file and tell me what to cut, and what should move into separate files that you only open when a task needs them?
6. Four rules: paste these into your agents file
Once the file is trimmed, add these four rules. The first three are copied from my own file, and the fourth is the rule Hermes wrote when I asked it to.
1. Prefer CLIs over MCP tools when both do the job (gh over the GitHub MCP, gws for Google Workspace). A CLI returns only the lines asked for.
2. Move heavy exploration (many-file reads, broad searches) to subagents. Keep only the summary in the main window.
3. Always talk in ASD-STE100 simplified technical English and say only what needs to be said. Report only the elements needed for me to make the right decisions, explained clearly.
4. All sub-agents use the fastest, cheapest model available.
CLIs are built for machines to call, so they run faster and use fewer tokens than the matching MCP.
Rule 2 keeps file exploration out of your main chat. The sub-agent does the reading and sends back a summary, so the main window carries the summary instead of every file.
Rule 3 is the one you'll notice first. If one-line prompts get you pages of reply, this cuts them right down and saves output tokens on every answer.
For rule 4, Haiku 5.5 is fast and cheap, which makes it a good sub-agent model. Claude Code already does something like this on its own, but it's still worth writing into the file.
Can you add these four rules to my agents file? [paste the rules]
7. Model and effort: match the thinking to the job
Of all seven, this is the one to keep in mind on every task. Model choice and effort level decide how much each job costs.
Hermes runs open models like Kimi K3, MiniMax M3 and DeepSeek V4 Pro through OpenRouter, and you pay for what you use. You can also connect a ChatGPT or Codex subscription, or an Anthropic one, and spend the allowance you already pay for.
Either way, pick the model per task. In Claude Code my default is Sonnet 5.5 on medium effort. In Hermes I step up to Kimi K3 or GPT-6 Astra, still on medium.
For a full client build, like the tool I'm making for an exam centre in the UK right now, I use GPT-6 Astra on extra high. I almost never go to max or ultra.
Effort is thinking time, meaning how much reasoning happens before you get a reply. Put a short, simple ask on high effort and it spends more tokens, takes longer and hands back the same answer.
How much you'll actually save
I can't promise you 95%. That's what the right jobs can reach, and your number depends on how much weight your setup carries now.
But if you've never audited your agent, and you've been adding rules, MCPs, skills and files that load on the first prompt, there's a lot to cut. My own overhead more than halved from the audit.
Every step above is in the video, including the dashboard settings I switched on. Watch it here: https://youtu.be/K7rooUY4-FQ
Book a free AI audit call
If you know AI could save your business time and money but you don't know what to automate first, book a free 30-minute call with me.
I'll go through your current processes and bottlenecks with you and find the repetitive work AI could take off your team's plate. Then I'll pick the one automation with the highest return to build first. I'll map out how it gets built, whether I build it for you or you build it with me on one-to-one calls.
You'll speak to me directly. Book your free AI audit
Start with change 1 today. It takes one prompt, and the report shows you where your own overhead is going.