⭐ Get the Agent OS + join 3,400+ founders inside the AI Profit Boardroom → Join AIPB ($69/mo)

GoldieBench Blog · 8 min read

Hermes Labs: What Nous Research's Experimental Plugins Do, How To Install Them, And The Models That Power Them

Hermes Labs explained: Nous Research's unsupported Hermes Agent plugins, how to install Inbox, Workflow Record and Telegram Business, and models to run.

Hermes Labs: What Nous Research's Experimental Plugins Do, How To Install Them, And The Models That Power Them — illustrated hero

Hermes Labs is Nous Research's set of experimental Hermes Agent plugins that Nous built but does not support.

The Hermes Labs repo lists three of them, called Inbox, Workflow Record and Telegram Business Mode.

They arrived in the plugin catalog with the Hermes Agent v0.22.4 release, dated 21 October 2026.

I'm Julian Goldie, and I haven't tested the Labs plugins on video yet, so this post is built from Nous Research's own release notes and READMEs.

I'll explain what Hermes Labs is, how to install and use each plugin and where the limits are, and then I'll show you how the models you'd pair with it score on GoldieBench.

Hermes Labs at a glance

This table is the whole of Hermes Labs on one screen.

PluginInstall commandWhat it addsWhat it does
Inboxhermes plugins install inboxThe hermes inbox commandIt shows everything in flight for the active profile in one read-only view.
Workflow Recordhermes plugins install workflow-recordThe hermes record command and a replay skillIt records a browser task once so /learn can turn it into a skill.
Telegram Business Modehermes plugins install hermes-telegram-businessA Telegram Business handler and /biz commandsIt drafts customer replies and sends nothing until you tap Send.

All three come from the NousResearch account on GitHub, and the repo is MIT licensed.

What is Hermes Labs?

Hermes Labs is a shelf of working plugins that Nous Research wrote and then chose not to maintain.

Nous says these plugins "started as Hermes features we decided not to put in the core."

Rather than delete that work, Nous published it so anyone who wants it can install it from the catalog.

The official docs site doesn't have a dedicated Hermes Labs page yet, so the repo README is where Nous defines the term.

QuestionWhat Nous Research says
Who built it?Nous Research built every Labs plugin.
Is it supported?No, and Nous makes no promise of fixes.
Can updates break it?Yes, because a core change may break a plugin before anyone notices.
Is it tested?CI runs every plugin against the current Hermes nightly, so breakage is at least visible.
Which catalog tier?The catalog lists Labs plugins as community, not official.
Can I own one?Yes, each plugin is up for adoption, and Nous will help you move one to your own repo.
Where are the supported plugins?They live in the separate hermes-official-plugins repo.

The v0.22.4 notes say the catalog gained "inbox and workflow-record from Hermes Labs," and that "Telegram Business moves to the Hermes Labs community tier."

A separate GitHub organisation called hermes-labs-ai also has a catalog plugin, and its entry lists the maintainer as independent of Nous Research.

That one is not part of Nous Research's Hermes Labs, so don't mix the two up.

How to install Hermes Labs

Update Hermes first, because the Labs entries arrived through the catalog in the v0.22.4 window.

A command-line install updates with hermes update, the Microsoft Store app updates through the Store, and Docker users pull nousresearch/hermes-agent:stable.

Then you install the plugins you want with the commands in the table above.

You can find any catalog plugin with hermes plugins search <term>, or browse the catalog inside Hermes Desktop.

To back out, hermes plugins disable my-plugin switches a plugin off and hermes plugins remove my-plugin uninstalls it.

How to use Hermes Inbox

Inbox adds one command, hermes inbox, and it adds no model tools, no hooks and nothing that runs during chats.

It reads your profile's process registry, cron jobs, active sessions, state.db and the shared kanban board, and it opens every database read-only.

SectionWhat shows up
Needs attentionStalled delegations, dropped results, failed or paused cron jobs, orphaned processes and blocked kanban tasks show up here.
In progressLive background processes, running delegations and running kanban tasks show up here.
Finished, result not yet deliveredCompleted delegations whose result hasn't reached its conversation show up here.
ScheduledEnabled cron jobs and their next five runs show up here.
Open surfacesLive CLI, TUI, Desktop and gateway chat sessions show up here with their titles.
hermes inbox --json prints the same snapshot as JSON for scripts and dashboards. hermes -p work inbox reads another profile, as long as the plugin is installed in that profile too.

Inbox can't show pending approvals or clarify prompts, because they live only in the memory of the process that asked.

How to use Workflow Record

Workflow Record adds hermes record, which watches a browser task you do once in Chrome, Chromium, Brave or Edge.

It attaches over the DevTools protocol, and the simplest way to start a reachable browser is /browser connect inside a Hermes chat.

You open the starting page, keep that tab in front and run hermes record --slug shop-login.

You do the task, press Ctrl-C, and Hermes saves a JSON recording and prints the exact /learn line to paste into a chat.

The bundled workflow-record:replay-recording skill tells /learn how to rebuild the task as a skill that replays with Hermes's browser tools.

Password, one-time-code and card fields are masked as {SECRET:name} inside the page, and Hermes masks them again before anything is saved.

Recordings stay local in <HERMES_HOME>/recordings, and nothing is sent anywhere until you run /learn, which hands the recording to your active model.

hermes record --list lists recordings, and hermes record --manual lets you type the steps when you can't use CDP.

How to use Telegram Business Mode

Telegram Business Mode turns your Hermes Telegram bot into a secretary for your Telegram Business account.

Every customer message gets a drafted reply in your own DM, with Send, Edit and Discard buttons.

The README says no auto-send exists, so nothing reaches a customer until you tap Send.

You need a Telegram Business subscription, Business Mode switched on for your bot in @BotFather, and Hermes v0.18 or later.

After installing it, you run hermes gateway restart, add the bot under Telegram's Settings, Business and Chatbots, and message it /biz.

/biz pause and /biz resume act as a kill switch, and /biz off <chat_id> mutes one customer chat.

Drafting uses your active Hermes model through the plugin LLM surface, so there's no separate key.

The limits of Hermes Labs

PluginLimit Nous documents
All threeNobody promises fixes, so a Hermes update can break them.
InboxIt doesn't show pending approvals or clarify prompts.
Workflow RecordIt records one tab only, and pop-ups or new tabs stop the recording.
Workflow RecordOnly password-like fields are masked, so an API key in a plain text field is recorded as typed.
Workflow RecordIt skips scrolling, hovering, drag and drop, file uploads and iframe navigations.
Workflow RecordIt works with Chromium browsers only, and Firefox or Safari need --manual.
Telegram Business ModeVersion 1 is text only and drafts each message without conversation history.

I'd re-run each Labs plugin after every Hermes update before trusting it again.

Hermes Labs vs official Hermes plugins

The easiest way to understand Hermes Labs is to compare it with the official plugins Nous does support.

QuestionHermes Labs pluginsOfficial Hermes plugins
Who builds them?Nous Research builds them.Nous Research builds them.
Who fixes bugs?Nobody promises fixes.Nous Research supports them.
Where is the code?It lives in NousResearch/hermes-labs-plugins.It lives in NousResearch/hermes-official-plugins.
Which catalog tier?They show as community plugins.They show as official plugins.
How do they ship?The release is a catalog pin in hermes-agent.They also install from the reviewed catalog.

Every catalog entry is pinned to a reviewed commit, so installing a Labs plugin still gets you the exact code Nous reviewed for the catalog.

What you don't get is anyone on the hook when an update breaks it.

A simple first week with Hermes Labs

On day one, I'd install Inbox and run hermes inbox every morning to see which jobs failed overnight.

Later that week, I'd pick one boring browser chore and record it with hermes record.

I'd review the JSON recording for anything sensitive before running /learn on it.

Then I'd test the new skill on a low-stakes run before letting it near anything important.

I'd only set up Telegram Business Mode if customers already message me on Telegram, because it needs the most setup steps.

Does Hermes Labs ship its own model?

No, it doesn't, and that's the point where GoldieBench becomes useful.

Inbox calls no model at all, so the model you run makes no difference to it.

Workflow Record's /learn step and Telegram Business drafting both use whichever model is active in your Hermes profile.

So the quality of the skill that comes out of a recording, and the quality of every customer draft, depends on the brain you pick.

Which models to run with Hermes Labs

GoldieBench scores models on one-shot builds, not on turning a browser recording into a skill or drafting a Telegram reply.

So treat these numbers as a guide to general build quality, not a direct Hermes Labs score.

ModelGoldieBench averageWhy you'd pick it for Hermes Labs
Claude Opus 58.27It's the highest single model in this table, so it suits /learn skill writing when quality matters most.
Hermes MoA8.17It's the Hermes Mixture of Agents setup, which blends models inside Hermes itself.
GPT-5.6 Sol8.16It sits a whisker behind Opus 5 on the same build tasks.
Qwen 3.88.10It's the strongest open-weights family entry in this table.
Kimi K37.89It's a strong mid-table pick if you already use Kimi.
GLM-5.27.77It's another solid open-model option for routine drafting.
Claude Opus 5.57.57It's newer than Opus 5 but scores lower on these one-shot builds.
Qwable 5 27B Coder7.14It's the best local model on the board, so it suits drafts that must stay on your machine.

For Telegram Business drafting, where each reply is short, I'd lean towards a cheaper or local model and save the top models for /learn.

For private workflows, the local models board shows which models run on your own hardware.

If you're choosing a main model for Hermes more generally, my post on the best Hermes Agent model goes deeper.

Who Hermes Labs is for

Hermes Labs suits operators who are happy to run unsupported code and check it after updates.

Inbox suits almost anyone running more than one agent, because it only reads.

Workflow Record suits people with repeated browser chores, and Telegram Business Mode suits people who sell through Telegram.

If you're moving a Claude Code or Codex setup into Hermes at the same time, read my post on Hermes Import Agent.

Also On Our Network

If you run an SEO business and want to talk through where agents and plugins like these fit, you can book a free SEO strategy session with my team.

If you'd like a second pair of eyes on your whole agent stack, my team can review it in a free SEO strategy session too.

Install Inbox first, pick a strong model for /learn, and you'll get the most out of Hermes Labs.

FAQ

What is Hermes Labs?

Hermes Labs is Nous Research's set of experimental Hermes Agent plugins that Nous built but does not support. They began as Hermes features Nous chose not to put in the core, and the catalog lists them in the community tier.

Which plugins are in Hermes Labs?

There are three: Inbox, which adds hermes inbox; Workflow Record, which adds hermes record; and Telegram Business Mode, which drafts Telegram customer replies for your approval.

How do I install Hermes Labs?

Update Hermes, then run hermes plugins install inbox, hermes plugins install workflow-record or hermes plugins install hermes-telegram-business.

Does Hermes Labs come with its own AI model?

No. Inbox calls no model, and Workflow Record's /learn step and Telegram Business drafting use whichever model is active in your Hermes profile.

Which model should I use with Hermes Labs?

On GoldieBench one-shot builds, Claude Opus 5 averages 8.27, Hermes MoA 8.17 and GPT-5.6 Sol 8.16. Qwable 5 27B Coder averages 7.14 and is the best local option. GoldieBench measures builds, not skill writing, so treat it as a guide.

Is Hermes Labs supported by Nous Research?

No. Nous makes no promise of fixes, and a core change may break a plugin before anyone notices, although CI runs each plugin against the Hermes nightly.

The same stack Julian uses

Run this stack yourself.

Every demo on this bench was built inside the Agent Operating System — one prompt, one shot, single HTML file out. The Agent OS, the prompts, the templates, the weekly walkthroughs and 3,400+ founders shipping with it every day all live inside the AI Profit Boardroom.

3,400+founders
258documented wins
38countries
$69/momonthly