Hermes Agent Tokenjuice: Shrink Long Terminal Output Before the Model Reads It

Tokenjuice is a new plugin for Hermes Agent that shortens long terminal output before the AI model reads it. It joined the official Hermes plugin catalog on the 2nd of November 2026. Its maintainer is Vincent Koc, who builds the same tool for many other AI coding apps in the tokenjuice project on GitHub.

On the 4th of November, a post by HermesAgentTips on X shared a test on Hermes 0.22.6. A listing of 700 lines went from 44,082 characters to 774. A "git status" check went from 670 characters to 223. The account is an unofficial tips page for Hermes users.

Tokenjuice, in plain words

When Hermes runs a command on your computer or server, the command prints text back. Hermes hands all of that text to the AI model, so the model can decide what to do next. The model reads it in tokens, small pieces of text, and most providers bill by the token.

Some commands print a lot. A folder listing, a test run or an install log can fill thousands of lines, and most of them say nothing new.

• It trims on your machine. Tokenjuice runs locally. According to the catalog page, the plugin sends nothing over the network and needs no account or key.
• It keeps the useful part. On a long listing, the post on X says the model still sees the top and the bottom, plus a count of the lines that were skipped.
• It only touches commands that worked. A command that fails comes through in full, so the model still sees the whole error.
• It follows fixed rules. The project describes itself as a deterministic compactor: the same output always gets the same short version, by written rules, with no AI guessing.

In the test above, 774 characters out of 44,082 is under 2% of the original. Your own saving depends on which commands your agent runs.

What it costs and what you need

Tokenjuice is free and open source under the MIT licence.

• Hermes Agent version 0.22.5 or newer. The current release is 0.22.7.
• Node.js version 22 or newer. Node.js is a common helper program that runs JavaScript tools. The plugin calls a small Tokenjuice program, and that program needs Node.js.
• A terminal on the machine where Hermes runs. A terminal is the text window where you type commands.
• Linux, macOS or Windows. The catalog lists all three.

Step by step: install Tokenjuice

Step 1. Read the catalog entry first. Tokenjuice is a community plugin, which means Nous Research reviewed its listing but did not write it. This shows its details:

hermes plugins info tokenjuice

Step 2. Install the Tokenjuice program itself. The plugin is only a bridge, and this is the part that does the trimming:

npm install --global tokenjuice

npm comes with Node.js. On the 4th of November 2026 this installs version 0.10.0, the same version as the plugin.

Step 3. Install the plugin from the catalog:

hermes plugins install tokenjuice

Hermes installs the exact copy that the catalog reviewed, fixed to one commit. Updates reach you only after the catalog reviews them again. Hermes then asks "Enable 'tokenjuice' now?". The default answer is no.

Step 4. If you said no, switch it on now:

hermes plugins enable tokenjuice

Step 5. Check that everything is found:

hermes plugins doctor tokenjuice

If the check cannot find the Tokenjuice program, it is not on your system's path, the list of folders where the computer looks for programs. Install it again from Step 2, or point Hermes to it with a setting called TOKENJUICE_BIN, set to the full location of the program.

Step 6. Start a new chat and let Hermes do its normal work. Nothing else changes. You keep talking to Hermes the same way.

Important: the catalog page also has an "Open in Hermes Desktop" button. It opens the same install inside the Desktop app, where you confirm each part before anything installs. You still need Step 2 on that machine.

When it helps most

• Agents that work in folders and code. Listings, searches, "git status", test runs and build logs are the long, repetitive outputs Tokenjuice is made for.
• Long jobs that run many commands. Each command's printout stays in the chat history, so every later step pays for it again. Shorter printouts keep the whole conversation smaller.
• Small or local models. A model with a short memory, its context window, fills up fast. Less noise leaves more room for the actual task.

It helps less if your agent mostly chats, writes emails or browses the web, because those jobs run few terminal commands.

What it leaves alone

The plugin page lists the cases where Hermes keeps the original, full output:

• Commands that fail.
• Commands running in the background.
• Exact file reads, such as "cat", "head" or "tail". The project's rules keep these whole, however long the file is, because the model often needs every line.
• Requests for the raw output.
• Output over 2 MB, or anything Tokenjuice cannot finish within five seconds.
• Any case where the Tokenjuice program is missing or answers with something unexpected.

Hermes also keeps the command's exit status, the signal that says whether it worked, and still applies its own size limits and hiding of secrets afterwards, according to the project's Hermes guide.

When something looks off: what to check

• The model says a list is cut short. That is the plugin working. Ask Hermes to read the file directly, or switch the plugin off for that job.
• Nothing seems shorter. Run the doctor check from Step 5. Short outputs, failed commands and file reads are left as they are by design.
• It stopped after a Hermes update. Run the doctor check again. The plugin needs Hermes 0.22.5 or newer.
• You want everything back the way it was. Turn it off, or remove it:

hermes plugins disable tokenjuice
hermes plugins remove tokenjuice

Is it safe to run?

The catalog page spells out what the plugin does: it runs the Tokenjuice program on the output of every successful foreground command and replaces what the model sees with the shorter version. It never runs the command it is given a second time, it changes no Hermes permissions and it sends nothing over the network.

The Hermes documentation adds one rule worth remembering for every plugin: a catalog listing means the entry was reviewed when it joined, which is different from a full security audit. That is why Step 1 comes first.

Quick answers

Does Tokenjuice change what my agent does? No. It only changes what the model reads after a command has run. The command itself runs as normal.

Will I save money? Fewer characters sent to the model means fewer tokens, and most providers bill by the token. How much depends on your commands. The 98% cut in the test above came from one very long listing.

Does it work with OpenClaw too? The tokenjuice project lists an integration for OpenClaw as well, with its own setup. This guide covers Hermes only.

Can I undo it? Yes, with one command: "hermes plugins disable tokenjuice".

OneClickClaw runs Hermes Agent on a dedicated server of your own, which stays online while your computer is off. Every plan is BYOK, so your AI provider bills you directly, and a lighter token load shows up on that bill. The Starter plan is EUR 14.99 a month, with a 7-day free trial.

Related guides

• OpenClaw Token Bill: Find the Leak and Cut Costs
• Hermes Agent's 18 Official Plugins Explained
• How to Install Hermes Agent and Start Your First Chat
• Hermes Agent broke after an update: what to check first

Sources: the Tokenjuice entry in the official Hermes Agent plugin catalog and the Hermes plugin documentation at hermes-agent.nousresearch.com, the tokenjuice project on GitHub and its npm package, all read on the 4th of November 2026, and the HermesAgentTips post on X of the 4th of November 2026.