A promise at the edge of the group chat.
Hermes AgentTwenty promises.
All checked.
Nous Research posted twenty things you should be able to do with your agent. We read the documentation to see which ones hold.
Hermes Agent is Nous Research’s open-source AI agent: MIT licensed, run on your own machine or a server, with memory that persists between sessions and skills it writes for itself. Checked against its own documentation, 15 of the 20 promises in its 5 October post are documented, 4 are partly documented and 1 is a promise rather than a feature.
An original folkfox special. 6 October 2026.
A promise is a map.
The documentation is the ground.
Chapters
Chapter one. The invitation.
Suppose your assistant lives in the family group chat.
Picture a Saturday that has not started yet. In the family group chat, a message lands before anyone is up. The aquarium opens at ten. The nap window is one o’clock. There is a café four minutes away, and it has a high chair. This is a hypothetical scene, not a screenshot. It is built from what two people actually posted this week.
On the evening of 5 October, Sergey Kaplich wrote that he had set up a Hermes agent for his family, with a separate profile holding the essentials about each member and a Telegram group chat shared with his wife. The next morning, witcheer described the same Hermes Agent Telegram set-up, with the same examples: the aquarium, the nap, the rainy afternoon. Both are self-reports, and the second repeats the first. Neither is a test.
That is the pull of the thing, and it is real. An assistant that remembers the nap time is a different animal from a chat window that forgets you between sessions. The question this page asks is narrower, and more useful: what, exactly, is it allowed to remember, do and say, and who decided?
Look a little closer
We could not verify who either poster is, or how their set-up is configured, and we have not tried. They are quoted for what they say, not as evidence that the arrangement works for anyone else. The job-search story in chapter eight gets the same treatment.
- Family agent
Good morning. The aquarium opens at ten, the nap window is one o'clock, and there is a café four minutes away with a high chair.
Illustrative scene. No model runs here, and your three choices stay on this page.
Chapter two. The manifesto.
Twenty sentences, and one signature line.
At 22:43 Malta time on 5 October, Nous Research posted a list. Twenty lines, each beginning “You should be able to”, each about an agent: smarter models, cheaper models, local models, memory you control, a machine that never touches the internet, code you can read and change. The last line is the signature. “Your agent should be yours.” When we read it on 6 October the post showed about 122,000 views and 3,300 likes.
Attention is not evidence. A list this good is easy to cheer, because everybody wants it. A fox does not cheer a list. A fox follows each line to the ground and checks whether anything has stood there.
The summary that did the rounds on 6 October told the week another way, as a major update with desktop, voice and plugin upgrades. We will come back to that headline in chapter nine.
Our method
For each line we looked for a sentence in Nous’s own documentation, repository or pricing page that says the thing can be done. Documented means a page says so. It does not mean we tested it, and it does not mean it works well. Partly means a page covers some of the line. A promise means no page could settle it. That makes this page an audit of promises, not a hands-on Hermes Agent review.
Optional. Lock a guess now and chapter three will answer it.
Chapter three. The audit.
Guess first. Then open the cards.
Each card holds one promise, word for word. Choose what you think the documentation says, then open the card. There is no penalty, and nothing leaves this page.
The verdicts are ours, and you can disagree. Every card links the sentence we relied on, so you can check our reading against the page itself.
A note on “documented”
Documentation moves faster than anyone’s reading of it. The project tagged six releases between 31 August and 24 September. Treat each verdict as dated 6 October 2026.
0 of 20 opened.
- Documented 0
- Partly 0
- A promise 0
The brain
Models: which one, where, how hard it thinks.
01
“You should be able to use smarter models in your agent.”
Documented
The Portal documentation advertises more than three hundred frontier models behind one subscription, and you can bring your own provider instead.
“300+ frontier models, one bill”
Nous Portal docs
The catch. Nous's pages disagree on the count: 300+ in the docs, 200+ on its subscription page.
02
“You should be able to use cheaper models in your agent.”
Documented
The subscription page lists a free tier that offers free models only, and Hermes accepts any provider you configure.
“Free models only”
Nous Portal pricing
The catch. Cheaper is the model's price, not the agent's: auxiliary tasks can add cost on an expensive main model.
03
“You should be able to use local models in your agent.”
Documented
The FAQ lists Ollama, vLLM, llama.cpp and SGLang as supported local routes.
“Local models”
Hermes FAQ
The catch. A local model needs enough GPU memory, and the FAQ names a 64,000-token context minimum.
04
“You should be able to use a different model for every job.”
Partly documented
Each Bot or profile has its own model, and background tasks such as compression take their own provider and model.
“Each Bot has its own role, model, memory, skills, and avatar”
Bot Mode docs
The catch. That is a model per Bot and per background task. The FAQ lists a section on multi-model workflows, but we read its heading and not its steps, so this is the likeliest card to move.
05
“You should be able to switch models in the middle of a conversation.”
Documented
The /model slash command switches between models you have already set up, without leaving the session.
“Switch between already-configured models without leaving a session”
CLI reference
The catch. It cannot add a new provider mid-chat. That still needs the full setup wizard.
The data
What it holds, forgets and reaches.
06
“You should be able to choose what data your agent has.”
Partly documented
Toolsets can be switched on and off per platform, and secrets live in a separate file. We found no single page that lists every data source an agent can reach.
“toolsets that can be enabled or disabled per platform”
Features overview
The catch. Choice here is assembled from several settings, not one data-permissions screen.
07
“You should be able to choose what it remembers and what it forgets.”
Documented
The agent adds, replaces or removes its own memory entries, and the memory files are plain files you can open.
“it can add, replace, or remove entries”
Memory docs
The catch. The agent saves automatically. You choose by editing afterwards, not by approving each note.
08
“You should be able to choose what computer your agent runs on.”
Documented
The README lists seven terminal backends: local, Docker, SSH, Singularity, Modal, Daytona and Vercel Sandbox.
“Seven terminal backends”
Hermes README
The catch. The default local backend runs commands on your machine with no isolation.
09
“You should be able to run it on a machine that never touches the internet.”
Documented
The FAQ answers yes to running offline with a local model, configured through a custom endpoint.
“Can I use it offline / with local models?”
Hermes FAQ
The catch. Chat apps such as Telegram are internet services, so an offline agent is a terminal agent.
The home
Where it lives and how you reach it.
10
“You should be able to run it while your laptop is closed.”
Documented
The README says it is not tied to your laptop: talk to it from Telegram while it works on a cloud VM.
“It's not tied to your laptop”
Hermes README
The catch. Closed means elsewhere. A sleeping laptop is not an always-on service.
11
“You should be able to reach it from the apps you already use.”
Documented
The gateway connects Telegram, Discord, Slack, WhatsApp, Signal, SMS, Email, Teams and more from one background process.
“Chat with Hermes from Telegram, Discord, Slack, WhatsApp, Signal”
Messaging gateway docs
The catch. Not every platform supports every feature. The gateway docs carry a table of what each one can do.
12
“You should be able to choose how your agent thinks.”
Documented
Thinking level is a setting: none, minimal, low, medium, high, xhigh, max or ultra, per task and as a default.
“Thinking level for that task's LLM calls”
Configuration docs
The catch. More thinking costs more tokens. The setting is yours, and so is the bill.
The mind
What it can read, see and learn.
13
“You should be able to read the prompt it runs on.”
Partly documented
SOUL.md is the first thing in the system prompt and you can edit it. The docs list the rest of the prompt stack, but we did not find a page that prints the assembled prompt.
“SOUL.md is the primary identity”
Personality docs
The catch. You can read and change the identity file. Reading every assembled prompt is a different promise.
14
“You should be able to see every tool it has, and turn off the ones you don't want.”
Documented
Tools are grouped into toolsets that can be enabled or disabled per platform, and hermes tools configures them.
“Configure enabled tools per platform”
CLI reference
The catch. Plugins and skills add tools later. Check the list again after you install any.
15
“You should be able to teach it something once and never explain it again.”
Documented
The /learn command turns something you know, or a pile of reference material, into a reusable skill.
“turn something you already know”
Skills docs
The catch. Never is strong. A skill loads when relevant, and the agent still has to choose to use it.
16
“You should be able to know what it did, and when, and why.”
Partly documented
hermes sessions browses and exports every session, and hermes logs tails the log files. That answers what and when.
“Browse, export, prune, rename, and delete sessions”
CLI reference
The catch. Why is the hard part. A log shows actions; it does not explain the model's reasons.
The keys
Ownership, in the end.
17
“You should be able to export your agent.”
Documented
hermes backup zips the whole Hermes home directory, and a profile can be exported on its own with credentials stripped.
“Back up Hermes home directory to a zip file”
CLI reference
The catch. A full backup includes your .env and auth.json, so treat the zip like a password. A profile export strips credentials, and a new machine may need re-authenticating.
18
“You should be able to read the code.”
Documented
The project is public on GitHub and described by Nous as free and open source under the MIT licence.
“free and open source under the MIT license”
Hermes home page
The catch. Readable is not audited. We have not reviewed the code, and we found no independent review.
19
“You should be able to change the code.”
Documented
The MIT licence is a permissive licence, and the project publishes a contributor guide and a plugin catalogue.
“free and open source under the MIT license”
Hermes home page
The catch. Changing the code means maintaining your fork against a project that tags a release every week or so.
20
“Your agent should be yours.”
A promise, not a feature
This line is the signature, and no page can settle it. The code and the notebook sit in your own folder. Most of the best models are rented.
“Hermes Agent is free and open source under the MIT license.”
Hermes home page
The catch. Own the harness and you can still rent the brain. Read chapter ten before you answer.
Chapter four. The notebook.
A small notebook, not a brain.
Say “persistent memory” and most people picture a hive mind that never forgets. Open the manual and you find something humbler.
Hermes keeps two small files. MEMORY.md holds the agent’s own notes: environment facts, conventions, things it learned. USER.md holds what it knows about you. The limits are in the documentation: about 2,200 characters for the first and 1,375 for the second. That is a few paragraphs, not a life.
The files load into the prompt at the start of every session, so everything in them is always in the agent’s head. The limit is the design. Character limits keep memory focused, and it does not auto-compact. The agent decides what to save, and it saves automatically, so you do not need to ask.
Everything else lives in a second place. Every session is stored in a local SQLite database with full-text search, so the agent can find something you discussed weeks ago even if it is not in the notebook. A notebook for what must always be known. An archive for what can be found.
Look a little closer
For deeper memory, Hermes ships plugins called memory providers. Where a provider stores the data is the first question to ask of each one. The promise to choose what it remembers is only as strong as the place the memory sits.
You are the agent's editor.
Keep what must always be known.
0 of 2,200 characters used
Left out of the notebook
Everything you leave out is still in session history, found by search when needed.
Illustrative family notes. The 2,200-character limit is from the memory documentation; the notes are ours.
Chapter five. The forge.
Teach it once.
Here is the promise people repeat most: teach it something once and never explain it again. The mechanism is a skill.
A skill is a written procedure the agent keeps in its skills folder. The agent can create, update and delete its own through a tool called skill_manage. When it works out a non-trivial workflow, it can save the approach for next time. Every installed skill also becomes a slash command.
You can teach it on purpose, too. The /learn command turns something you already know, or a pile of reference material, into a reusable skill without hand-writing the file. Skills follow the open agentskills.io standard and load only when relevant, so a long library does not clog the conversation.
Memory stores small durable facts. Skills store longer procedures. The documentation describes the two working together as the self-improvement loop.
Look a little closer
Skills installed from a hub pass a security scanner that checks for data exfiltration, prompt injection and destructive commands. A skill is also a way to hand an agent instructions, which is why that scanner exists.
What the agent would keep
Slash command:
A teaching sketch. Hermes writes and manages its own skill files, so use this to plan the lesson, then ask the agent to /learn it.
Chapter six. The address.
Where does it sleep?
A chatbot lives on someone else’s computer. A Hermes agent lives where you put it.
The README lists seven terminal backends: local, Docker, SSH, Singularity, Modal, Daytona and Vercel Sandbox. It says to run it on a five-dollar VPS, a GPU cluster or serverless infrastructure, and not to tie it to your laptop.
Where it lives decides what it can touch, what it costs and what happens when you shut the lid. The default local backend runs commands directly on your machine with no isolation. Docker adds hardening, and SSH keeps the work on another server.
On privacy, the FAQ is plain: API calls go only to the provider you configure, there is no telemetry, and conversations, memory and skills are stored locally in the .hermes folder. That is the vendor’s statement about its own software. The code is open, so it can be checked, but we have not audited it.
Money is the other address. Nous Portal pricing is a subscription ladder: Plus is $20 a month with $22 of monthly credits, and there is a free tier with free models only. The software stays free either way, and you can skip Portal and bring your own provider.
Choose one from each row to see what the documentation says about your sketch.
Chapter seven. The keys.
Who else can talk to it, and what can it do?
The same features that make an agent useful make it risky. It reads files, runs commands, browses the web and answers messages. The Hermes Agent security documentation is candid that this needs guardrails, and it lists them.
Before it runs a command, Hermes checks it against a list of dangerous patterns. The approval system has three modes. Smart, the default, asks a second model to assess the risk. Manual always asks you. Off disables the checks, the same as running in YOLO mode.
Strangers get a different door. An unknown person who messages the bot receives a one-time pairing code, and the owner approves it from the command line. Codes expire after an hour, and five failed approvals lock the door for an hour.
Context files are scanned for prompt injection and invisible Unicode characters before they reach the prompt. Credential stores such as .ssh and .aws are blocked outright. None of that makes an agent safe to point at anything. It makes the choices visible, which is the point of this chapter.
What the documentation does not promise
Container backends skip the dangerous-command checks because the container itself is the boundary. Keeping that boundary sound is your job, not the checker’s. We found no independent security audit of Hermes in the sources we read, and we did not run one.
A stranger messages the bot
- A stranger sends your bot a direct message.
- The bot replies with a one-time pairing code and nothing else.
- You approve the code from the command line, within the hour.
Press next to walk the door.
Illustrative controls. Nothing here runs Hermes or sends a message.
Chapter eight. The evidence trail.
Who is saying it, and how would they know?
You have read the promises and the manual. Now the fox leaves three clues on the path and asks you to name the kind of evidence each one is.
There is no score, and no verdict on Hermes waiting at the end. The point is the habit: before you repeat a claim about an agent, name who is making it.
Look a little closer
The job-search report is a single self-reported account, dated 20 September 2026. It credits a whole workflow, not only an agent, and it cannot separate what Hermes did from what the person, the market and a job-search tool did. The best that can be said is that it is a real person describing a real outcome.
The fox’s evidence trail
Clue 1 of 3Name who is speaking. Follow the clue only as far as it goes.
Choose a lantern to light the next part of the trail.
Chapter nine. The headline and the record.
Five claims from a summary, checked.
On 6 October a news-app summary of Hermes Agent was doing the rounds, the kind that fits a whole story into four lines. We read it as a compression, and checked each claim against the record.
Open each claim to see its stamp. Three hold in some form. One does not hold as stated, and one has turned a credit into a price.
A summary is only as good as the care taken in squeezing it. The habit worth keeping is to ask which sentence the summary came from.
“It lets users run powerful AI anywhere, from local machines to servers.”
Holds
The README lists seven terminal backends, from local to SSH, and says to run it on a five-dollar VPS, a GPU cluster or serverless infrastructure.
Source: Hermes README
“It supports over 300 models and platforms such as Telegram and Discord.”
Holds, with a caveat
The Portal docs say 300+ frontier models, but Nous's own subscription page says 200+. The gateway docs list Telegram, Discord, Slack, WhatsApp and more.
Source: Nous Portal docs
“A major 5 October update added desktop enhancements, voice features and better plugins.”
Not found as stated
The newest tagged release we could find is version 0.21.5 of 24 September, with full notes deferred to 0.22.0. That release carries the Desktop plugin SDK and voice shortcuts. The 5 October event we can date is the manifesto post.
Source: GitHub releases
“Real users report landing job interviews and organising family activities in group chats.”
Holds as self-reports
One GitHub issue reports 21 applications, 7 interview processes and 2 offers in 8 weeks using Hermes with a job-search tool. Two posts describe a family group chat. All three are one person's own account.
Source: GitHub issue, 20 September
“It is backed by a $22 monthly subscription option.”
Garbled
Plus is $20 a month, and it comes with $22 of monthly credits. Super is $100 with $110, and Ultra is $200 with $220. The core software is free.
Source: Nous Portal pricing
Chapter ten. The charter.
You can own the code and still rent the brain.
Return to the family group chat. Your three first instincts are waiting in the charter below, along with the choices you made on the way.
Here is what the audit found. Most of what Nous promises is written down. The lines that are only partly written down are the ones that matter most: what an agent can reach, what it quietly keeps, and how you would know what it did.
And “yours” has a boundary. The code is MIT licensed, and the notebook sits in your own folder. The models are mostly someone else’s, and the best are rented by the token or by the month. You can own the harness and still rent the brain.
So write the charter before you write the first prompt. It takes two minutes, and it is the one document that no vendor can write for you.
Your charter
A SOUL.md starter
SOUL.md is the agent’s identity file, loaded first in its prompt, per the Hermes documentation. This is a starter you edit, not a configuration we have tested.
Quick answers
What is Hermes Agent?
Hermes Agent is Nous Research's open-source AI agent. It is MIT licensed, runs on your own machine or a server, remembers between sessions, writes reusable skills and answers in apps such as Telegram and Discord.
Is Hermes Agent free, and what is Nous Portal pricing?
The software is free and open source under the MIT licence. You pay for the models it calls. Nous Portal pricing is a subscription option: Plus is $20 a month with $22 of monthly credits, and there is a free tier with free models only.
Where does Hermes Agent keep its memory?
Locally, in the .hermes folder. Two small files hold persistent memory, with limits of about 2,200 and 1,375 characters, and a local SQLite database holds searchable session history.
Can Hermes Agent run offline?
Yes, with a local model such as one served by Ollama or llama.cpp, according to the FAQ. Chat apps like Telegram still need a network, so an offline agent is a terminal agent.
Is it safe to put Hermes Agent in a family group chat?
That depends on the setup. Unknown senders get a one-time pairing code the owner must approve, and risky commands pass an approval check. The default local backend has no isolation, so many people choose Docker or a separate server.
What does 'your agent should be yours' mean?
It is the closing line of a twenty-line post Nous Research published on 5 October 2026. It promises control of models, data, memory, hardware and code. Our audit found fifteen of the twenty lines documented and one a promise rather than a feature.
Sources
- Nous Research published twenty 'you should be able to' lines on 5 October 2026, ending 'Your agent should be yours.'Primary: Nous Research on X
- Hermes Agent is free and open source under the MIT licence.Primary: Hermes home page
- Hermes lists seven terminal backends and can run on a $5 VPS or serverless infrastructure.Primary: project README
- Persistent memory is bounded: MEMORY.md at 2,200 characters and USER.md at 1,375, with session search in SQLite.Primary: memory documentation
- The agent creates, updates and deletes its own skills, and /learn turns knowledge into a skill.Primary: skills documentation
- Approval modes are smart, manual and off, and unknown users receive a one-time pairing code.Primary: security documentation
- API calls go only to the provider you configure, and there is no telemetry.Primary: FAQ
- The configuration page documents terminal backends and a per-task thinking level.Primary: configuration documentation
- The CLI reference documents hermes backup, hermes sessions, hermes logs, /model and hermes tools.Primary: CLI reference
- SOUL.md is the agent's primary identity, loaded first in the system prompt.Primary: personality documentation
- Each Bot is a profile with its own model, memory, skills and credentials.Primary: Bot Mode documentation
- The messaging gateway connects Telegram, Discord, Slack, WhatsApp, Signal and more from one process.Primary: messaging documentation
- Nous Portal offers 300+ frontier models and is the recommended way to run Hermes Agent.Primary: Nous Portal documentation
- Plus is $20 a month with $22 of monthly credits; Super $100 with $110; Ultra $200 with $220.Primary: Nous Portal pricing
- The newest tagged release found is v0.21.5 of 24 September 2026.Primary: GitHub releases
- A software engineer reported 21 serious applications, 7 interview processes and 2 offers in 8 weeks.Self-reported: GitHub issue
- Sergey Kaplich described a Hermes agent in a family Telegram group on 5 October 2026.Self-reported: X post
- witcheer described a family profile and Telegram group on 6 October 2026, repeating Kaplich's examples.Self-reported: X post
| Date | Public record | On this site |
|---|---|---|
| Version 0.21.0 tagged. | AI consultancy | |
| Version 0.21.1 tagged. | What did Dario see? | |
| Version 0.21.5 tagged: Desktop plugin SDK, Simple and Advanced modes, voice shortcuts. Full notes deferred to 0.22.0. | Who taught Claude to care? | |
| Nous Research posts twenty promises at 22:43 Malta time. | The newsroom |
What we could not check
We read documentation and did not test every feature, so a documented line is a claim by the maker, not a result. We found no independent benchmark of Hermes against other agents, and no independent security audit. We did not audit the code. The post’s view and like counts were read on 6 October and will have moved. Desktop, voice and Bot Mode features changed during the weeks we looked, so read the date on every verdict.















