Hermes Agent is a free, open-source AI agent developed by Nous Research. It runs on a server the user controls, operates continuously, and can be reached from a messaging app, a web dashboard, and a desktop application. This guide covers the setup process in six stages, from hosting the agent to running a team of agents on a shared task board. It requires no programming experience.

Key Concepts

What Hermes Is

Hermes is software that runs on a server. The user pays for two things:

• The AI model the agent uses to reason, billed per usage.
• The server the agent runs on, which costs a few dollars per month.

Hermes can use the same AI models offered by ChatGPT and Claude. Its difference lies in the features built around those models.

How Hermes Differs from ChatGPT and Claude

ChatGPT and Claude both run scheduled tasks, retain information about the user, read email, and send messages. Hermes differs in three areas that these services do not provide:

Two further differences apply:

• Memory structure. ChatGPT and Claude keep context within separate projects. Hermes uses one agent that learns the user's full working context and applies it to any task.
• Skills. Hermes records recurring procedures as skills, which are step-by-step instructions the agent reuses. Skills allow the agent to improve at specific work over time.

Hermes is in the same category as OpenClaw, an always-on personal agent. Users migrating from OpenClaw can ask their Hermes agent to import their OpenClaw configuration, and the agent guides the process.

Recommended Approach to Examples

The examples in this guide reflect the author's own work. Readers should adapt each prompt to their own job, study, or household tasks. Learning progresses faster when the work concerns a topic the user cares about.

Level 1: Hire the Agent and Give It a Home

The agent requires a machine that meets three conditions:

1. It stays awake. Scheduled jobs, monitoring, and outgoing messages depend on the machine running at all hours. A laptop in a bag does not meet this requirement.
2. It is separate from personal data. The agent runs commands, installs software, and writes and deletes files. The machine should hold no personal photos, passwords, sensitive work documents, or browsers logged into financial accounts.
3. It is reachable. The phone, laptop, and tablet connect to the same agent.

A virtual private server (VPS) meets all three conditions. A VPS is a computer in a data center that runs continuously and accepts logins only from its owner. The smallest plans cost a few dollars per month.

Installing Hermes on a personal computer is possible, but the agent then runs only while that computer is awake and operates alongside personal files. A rented server is the recommended option for a first agent.

Most hosting providers require installing Hermes manually through a terminal. This guide uses Hostinger, which offers a one-click Hermes template. The template installs Hermes, places it behind a secure web address, provides a login page, and includes backups, one-click updates, and a restart button.

Interface labels, buttons, and features change over time. When a label differs from this guide, follow the general sequence of steps.

Step 1: Purchase the VPS

1. Open the Hostinger VPS plan selection page and click Choose plan.
2. Select the KVM 1 plan. It is the smallest plan and runs one agent comfortably. The plan can be upgraded later without reinstalling.
3. In the cart, select a registration period. Longer terms cost less per month, but the available discount requires at least 12 months. The shortest recommended term is 12 months.
4. Hostinger offers a 30-day money-back guarantee for any term.
5. Untick the pre-selected option labeled "ready to use AI." It enrolls the account in a separate AI service that this setup does not require. The AI model is connected in Level 1, Step 4.
6. Leave the other add-ons unselected.
7. Confirm that the discount appears on the page, then click Continue.
8. Register with Google, GitHub, or an email address, create a password, and optionally enter a phone number. Click Register.
9. Enter billing and payment details to complete the transaction.
10. After payment, Hostinger displays a confirmation page and routes the user to the Hermes setup form.

Step 2: Create the Admin Credentials

1. In the setup form, keep the admin username hermes or replace it with a custom username. Record the username.
2. Click the eye icon to reveal the admin password. Copy it.
3. Store the password in a password manager. Do not store it in a plain text file or any location others can access. This password controls access to the agent.
4. Click Deploy.
5. Wait for the application status to change from deploying to a green check mark with the status Running. This process takes a few minutes.

If the user lands on a bare server dashboard instead of the Hermes setup flow, complete these steps:

1. In the left menu, select VPS.
2. If no VPS appears, click Get VPS. If a VPS appears, click Manage.
3. Select Docker Manager in the left menu.
4. Select Catalogue.
5. Search for hermes, select Hermes Agent, and click Select.
6. Complete the onboarding form.

The Docker Manager application section shows two installed components:

• Hermes Agent, the agent itself.
• Traefik, which secures the connection to the server.

Step 3: Open the Web Dashboard

Setting up Hermes previously required a terminal and typed commands. In this configuration, the agent is reachable through a browser.

1. In Docker Manager, locate the Hermes Agent application.
2. In the Access column, click the open link. This opens the Hermes dashboard login page.
3. Bookmark the URL.
4. Sign in with the saved username and password.

The dashboard includes the following sections in the left navigation menu:

Selecting Chat opens a terminal-style chat window. A test message at this stage returns the error "No inference provider configured." An error also appears in the upper right corner. The agent is installed and running, but it has no model to supply reasoning. The gateway, the component connecting the agent to external messaging, also has nothing to connect to at this stage. The gateway is configured in Level 2.

Step 4: Connect a Model Provider

A model is the AI that performs reasoning. Hermes does not include a model. The user connects an account that supplies one.

This guide uses Nous Portal, which is built into Hermes and is operated by Nous Research. Connecting it requires logging in and does not require copying access codes.

Nous Portal terms referenced in this guide:

• Access to more than 200 models from major AI companies.
• Tools for the agent, including web search, webpage reading, image generation, browsing, and speech.
• A subscription of $20 per month, which provides$22 of usage credits. Unused credits roll over to the following month.

Setup procedure:

1. Create a Nous Portal account through the invitation link. Choose Google, email, GitHub, Discord, or another listed option.
2. Enter the six-digit verification code sent to the email address.
3. Complete checkout. The confirmation page shows the account balance loaded with credits.
4. In the Nous Portal account menu, open Billing to view the current subscription. Subscription tiers can be changed at any time.
5. In Hermes, select Keys in the left menu.
6. Next to Nous Portal, click Login.
7. In the authorization window, click Connect to allow Hermes to use the Nous Portal plan.
8. Return to the Hermes dashboard. Nous Portal now appears as connected on the Keys page.

Selecting a model:

1. Select Models in the left menu.
2. Click Change next to Main model.
3. Select a fast, low-cost model for early use, because learning involves many messages. The guide uses DeepSeek V4 Flash. It is strong at tool use and costs a fraction of a cent per message.
4. Model names change over time. The newest version of a model family can replace the version named in this guide, such as DeepSeek V4.1 or later releases. The version number is less important than the model's category.
5. Click Switch, then click Reload when prompted.

Testing the connection:

1. Open Chat in the upper left.
2. Confirm that the upper right shows the selected model.
3. Send Hello. The agent responds that it can help.
4. Send What tools do you have? The agent lists its built-in tools: a browser, terminal commands, file handling, code execution, web search, delegation to subagents, memory, and scheduling.

Alternative providers:

• Credit-based access through a service such as OpenRouter.
• A ChatGPT subscription. An agent consumes plan limits quickly.
• A Claude subscription. Anthropic's terms restrict these plans to Anthropic's own applications, so this guide does not cover that method. Nous Portal provides access to every Claude model on a per-message basis.

Additional providers can be added later from the Keys page or by asking the agent.

Level 1 Checkpoint

The agent has a server and a model. Before continuing, send the agent three questions about a personal topic and observe its responses without additional setup.

Level 2: Reach the Agent from Anywhere

The browser dashboard works from any location, but it requires opening a browser tab. This level connects the agent to Telegram, so it can reach the user from a phone, and installs the desktop application.

Telegram is the simplest platform to configure. Setup requires scanning a QR code and does not require copying codes. Telegram supports fast responses and combines text, voice notes, files, and photos in one thread. Hermes supports more than 20 other platforms, including Discord, Slack, and WhatsApp, which can be added later from the same page.

Step 1: Enable the Telegram Channel

1. Install Telegram on a phone or desktop if it is not already installed.
2. In the Hermes dashboard, select Channels in the left menu.
3. Locate the Telegram channel. It appears as disabled.
4. Click the toggle in the upper right corner to enable it.
5. Click Restart. A gateway restart notification appears.
6. Wait several minutes for the gateway to restart. If the gateway does not show as running after several minutes, refresh the browser.

After the restart, the gateway displays as running. The Telegram channel shows that no bot token is configured.

Step 2: Create and Connect the Bot

1. Click Create with QR to connect a Telegram bot.
2. Scan the QR code with the phone's camera app and open the link.
3. Telegram opens a conversation with the bot. Click Create bot, then Create.
4. The conversation opens with the agent, but the bot is not fully connected. Return to the Telegram chat list and select the bot conversation. If the conversation is missing, scan the QR code again.
5. The conversation displays the message "Connect this bot to your Hermes setup." Click Connect this bot.
6. The bot connection notification appears in Telegram. The bot also appears in the web dashboard, with its name and the allowed user, which is the user's Telegram ID.
7. Click Save and restart to lock in the configuration. The restart takes several minutes. Refresh the browser if the status does not update.
8. The Telegram channel and the gateway both display as connected and running.

Step 3: Start the Conversation

1. In Telegram, tap the bot username, which begins with @hermes_.
2. Tap Start. The start command does not begin the conversation, so a message must be sent.
3. Send a test message, such as Hello! Can you hear me? The agent responds.
4. The agent reports that no home channel is set for Telegram. The home channel receives results from scheduled jobs and cross-platform messages. Send /sethome to set Telegram as the home channel. Scheduled notifications then arrive in Telegram.

Security: The setup places the user's Telegram account on the bot's allow list. Other users who find the bot receive no response. Telegram bot chats are not end-to-end encrypted. Passwords and banking details must not be sent through the bot. The tasks in this guide do not require sensitive data in chat.

Step 4: Manage Threads Across Devices

The phone conversation and the dashboard conversation are separate threads. The agent has one shared memory, but it maintains multiple threads, similar to multiple email threads with one assistant.

Handoff procedure:

1. In the dashboard chat, start a conversation. For example, ask for a synopsis of The Godfather.
2. Type /handoff, add a space, and select Telegram from the list.
3. Press Enter. Telegram displays "Working here now" and summarizes the thread.
4. Continue the conversation in Telegram. The desktop version of Telegram shows the same thread.

If a handoff fails, ask the bot in Telegram to continue the earlier work. The agent searches its own history.

Useful slash commands:

Commands that begin with / work in all three access points: the desktop app, the browser dashboard, and Telegram.

Session management: Conversations in the same thread share context. Each message resends that context to the model, which increases token usage. When a new topic starts, a new session prevents unnecessary cost.

Step 5: Troubleshoot a Silent Bot

Check the following three items in order:

1. Credits and rate limits with the model provider. This is the usual cause. Hermes reports the condition in the last message.
2. The gateway. Confirm that the status shows as running. Click Restart if the status is abnormal. This fixes a silent bot in most cases.
3. The logs. Open Logs in the left sidebar. Filter by type: debug, info, warning, or error. Filter for errors when a problem occurs. Reading every line is not necessary.

If none of these steps resolves the problem, restart the application from Hostinger's Docker Manager. Click the three dots next to the application, then click Restart.

Step 6: Install the Desktop Application

The desktop app resembles the ChatGPT or Claude desktop applications. It keeps conversations in a sidebar, accepts file drops, supports voice input, and reads responses aloud. The desktop app connects to the server, so it uses the same model and memory as the browser and Telegram.

macOS installation:

1. Open the official download page at hermes-agent.nousresearch.com.
2. Select the download for macOS. Save the installer.
3. Open the installer and drag the Hermes icon into the Applications folder.
4. Open Applications and launch Hermes. macOS asks for confirmation that the app was downloaded from the internet. Click Open.
5. Click Install Hermes. The setup steps run automatically and may take several minutes.
6. Click to launch Hermes after installation.

The Windows procedure is the same, with the Control key replacing Command in keyboard shortcuts.

Provider screen: Skip the model provider screen by selecting I'll choose a provider later. Choosing a provider here creates a second agent on the local computer. The agent already runs on the server.

Connecting to the server:

1. Press Command + , (Control + , on Windows), or click the gear icon in the upper right corner.
2. In the left menu, select Gateways.
3. Under connection mode, select Remote gateway.
4. In the remote URL field, enter the server address. Copy the address from the browser dashboard's address bar, including everything before the / that precedes the page path. For example, if the address bar shows /chat after the domain, copy the domain only.
5. Paste the URL into the desktop app's remote URL field.
6. Click Sign in to authorize the desktop app. Enter the saved username and password, then sign in.
7. When the app displays "You're signed in," click Save and reconnect.
8. Close the settings page and open the chat.

Test: Ask the agent, What was my first ever message to you on Telegram? The agent returns the first Telegram message. The user now has one agent with three access points: the browser dashboard, Telegram, and the desktop app. All three share the same memory.

Version alerts: The message "Backend out of date" appears when the app is newer than the server. The alert can be dismissed or resolved by selecting Update Hermes.

Step 7: Use the Heads-Up Display and Voice Features

Heads-up display (HUD):

• Press Command + Shift + H to open a floating bar that sits above other windows.
• The bar's position signals to Hermes what content is under it, such as a specific page or application.
• The bar can be resized and moved anywhere on the screen.
• Exit HUD mode with the Exit HUD mode button, or press Command + Shift + H again to return to the main application window.

Voice features in the desktop app:

Level 2 Checkpoint

The agent has a chat on the phone, the desktop, and a voice interface. Before continuing, send a message from the phone while away from the computer, then continue the same thread in the desktop app.

Level 3: Assign Work to the Agent

Consumer chatbots can search the web, read files, and connect to accounts. Hermes differs in two respects:

1. Location of work. The agent has its own computer. Files it creates remain on the server, where it can retrieve them later and where scheduled jobs can write output.
2. Autonomy. The agent's operations are designed to run autonomously and improve with each run.

Step 1: Run a Research Task

1. In the desktop app, enter the request: Research the top three standing desks. Give me a one paragraph comparison with links.
2. The agent loads skills and uses tools to read the web. The interface shows the active tool, such as reading a specific news site.
3. The agent returns a written comparison with citations, rather than a list of links.

Step 2: Observe Permission Controls

1. Request the creation of three test files. The agent writes them to the /opt/data folder on the server. The files are a plain text file, a Markdown file, and a JSON file.
2. Request the deletion of the same three files. When permission is granted explicitly, the agent deletes them without further prompts.

When the agent decides on its own to delete or run a risky command during background work, it requests approval first. A prompt labeled Run appears in the desktop app. In Telegram, a popup appears similar to the model selector. Approval can be granted for the current session or for one action only.

Routine tasks, such as research, do not trigger prompts. Deleting files and running risky code commands do. Nothing executes until the user approves.

Progressive trust: Access should be granted in stages. Start with limited permissions, observe the agent's behavior, and expand access only when the user is comfortable. Accounts the agent uses should be separate from the user's primary accounts. Main email and any account with financial access should not be connected until the agent has demonstrated reliability.

Step 3: Process Files

In the desktop app: Drag a file into the chat, such as a PDF, spreadsheet, or screenshot.

In Telegram: Send a document or photo.

Example procedure:

1. Drag an AI index report PDF into the desktop app.
2. Enter: Summarise this and pull out the three numbers I should care about. The agent returns three figures.
3. Enter: Turn that into a short document I can send to a colleague. The agent produces a Markdown file, named AI index brief, and offers a preview or download option.
4. Download the file to the local Downloads folder.

File limits: Files sent through Telegram have a size limit. The desktop app supports larger files. Files the agent creates remain on the server. The agent and the user do not share a file system by default.

Step 4: Connect a Shared Google Drive

The shared folder solves the file system gap. Google Drive is used in this guide because many users already have an account. The same connection grants access to Google Calendar, Google Sheets, and other Google Workspace tools used in Levels 5 and 6.

Account setup:

1. Create a separate Google account for the agent. Do not use the primary account.
2. Share individual folders from the primary account into the agent's account. The agent sees only the folders the user shares.

Connection procedure:

1. In a new session, enter: Set up the Google Workspace skills so you can use my Google Drive, Calendar, and Sheets.
2. The agent proposes a scope that includes Drive, Calendar, and Sheets. It also adds Docs. It asks whether to include Gmail and Contacts. Gmail access is sensitive, and the user should include it only when comfortable with the connection.
3. The agent asks whether the account uses hardware security keys under Advanced Protection. Answer in plain text, for example: Regarding the scope, I'd like to use all those tools. My account does not include any advanced protection such as hardware or security keys.

Note on the clarification tool: Hermes normally uses a tool named Clarify to ask follow-up questions. In the recording, the questions did not load. A screenshot of the blank form was sent to the agent with the message: I can't see the questions. They're just loading forever. The agent then asked the questions in plain text, which worked. Users who see a similar failure can send a screenshot and describe the problem.

Google Cloud setup:

1. The agent provides a link to the Google Cloud Platform. Open it in a standard browser. The agent can browse with a snapshot of the default browser profile, which carries over existing cookies, but the guide opens the link manually.
2. Click Create project. Name the project Hermes Agent. Click Create.
3. Confirm that the top bar shows the project name.
4. Open APIs and services, then Enabled APIs and services. Click Enable APIs and services.
5. Enable the following APIs, searching for each one and selecting Enable:
   • Google Drive API
   • Gmail API
   • Google Calendar API
   • Google Docs API
   • Google Sheets API
   • Google People API (the modern version of Contacts)
6. Each API page displays Disable API when enabled. The button stops the API if clicked. Do not click it.

OAuth configuration:

1. In APIs and services, select Credentials.
2. Click Create credentials, then OAuth client ID.
3. If a yellow banner states that the consent screen must be configured first, follow these steps:
   1. Click Configure consent screen, then Get started.
   2. Enter the app name, Hermes.
   3. Select the user support email from the dropdown.
   4. Click Next.
   5. For audience, select External. Click Next.
   6. Enter the contact email address. Click Next.
   7. Check the box agreeing to the terms of service. Click Continue, then Create.
4. Return to Credentials, click Create credentials, then OAuth client ID.
5. Select Desktop app from the application type dropdown. Name it Hermes. Click Create.
6. Do not share the client ID or client secret. Download the JSON file and store it in an accessible location.
7. Drag the JSON file into the Hermes chat and send it.

Test user setup: If the app is in testing mode, the user must add their own email as a test user:

1. Open the link the agent provides to the audience page.
2. Locate the Test users section and click Add users.
3. Enter the email address and press Enter. Click Save.

Authorization:

1. The agent returns an authorization URL. Right-click the link and select Open in external browser.
2. Log in with the account the agent will use.
3. Google displays "Google hasn't verified this app." Because the user is the developer of this app, click Continue.
4. Select all scopes. These match the permissions Hermes requested. The scopes can be reduced before confirmation.
5. Click Continue. An error page appears. This is the expected result.
6. Copy the full URL from the browser address bar, return to Hermes, paste the URL into the message field, and send it.
7. Processing takes several minutes. The agent confirms that setup is complete and that all services are connected.

Troubleshooting: Take a screenshot of the error, send it to Hermes, and describe what appears. Another chatbot, such as ChatGPT or Claude, can assist if Hermes cannot resolve the issue.

Shared folder structure: After authorization, instruct the agent: Create a drive folder called Hermes with Inbox and Outbox inside and give me the link. The agent creates the folders and returns a link.

• Inbox: Files the user places for the agent to read.
• Outbox: Files the agent creates for the user, each accessible through a link.

The folder names and structure can be changed. The agent can also keep folders organized on request or on a schedule, filing items where they belong.

Step 5: Test the Shared Folder

Outbox test:

1. In the earlier summary conversation, enter: Save the summary you wrote as a document, upload it to Outbox in Google Drive, and send me a link that works from my phone.
2. The agent creates a Google Doc in the Outbox folder. Set sharing to Anyone with the link can view.
3. Open the link on a phone or in an external browser to confirm access.

Inbox test:

1. Upload a file, such as the State of AI in 2026 document, to the Inbox folder.
2. Enter the ambiguous request: Grab the newest file in my Inbox folder and give me a five bullet summary. The user does not specify the folder name in Google Drive.
3. The agent identifies the newest file in Inbox and returns a five-point summary.

Self-improvement: During the Outbox task, the agent wrote a document delivery skill, updated a skill for credential setup, and saved a memory entry. The user did not request these updates. The agent identified the process requirements and recorded them.

Step 6: Control Email Actions

Granting Gmail access raises the question of whether the agent sends messages on its own. The agent follows the instructions it receives, so control comes from the rules the user provides.

Recommended rule: Draft the message, never send it, and notify the user that it is ready. The user reviews the draft, edits it, and sends it personally. This process takes about 10 seconds of the user's time. No message is sent in the user's name without review.

After editing a draft, ask the agent to review the changes and retain them. Later drafts then require less editing.

Staged delegation:

1. Read and organize email for several weeks. Sorting, labeling, and extracting key items are the initial tasks.
2. Draft replies.
3. Send messages independently, if at all. This is the final stage.

The automated job in Level 5 never sends email. It reads confirmations and files them.

Step 7: Connect Other Tools with MCP

Many tools support the Model Context Protocol (MCP), a standard that allows an agent to read from and act within a tool. Hermes includes a catalog of ready-made MCP connections, and additional connections can be added by pasting a link.

Procedure:

1. In the desktop app, select Capabilities in the left menu.
2. Open the MCP tab. It displays the existing catalog of prebuilt connections.
3. Select a connection, such as Canva, and click Install.
4. Once installed, the connection appears in the list of servers. Select it, click Authenticate, and connect the account.
5. The agent can then work inside that account, in the same way as the Google Workspace setup.

To disable a connection without removing it, use the toggle. To remove it entirely, click the trash icon.

Possible MCP connections include Notion, GitHub, a customer database, or accounting software. The user should connect the tool, decide what the agent may change, and then assign work within that tool.

Step 8: Run a Calendar Task

Example request in a new session:

Look at my calendar for next week and add a 30-minute block called change studio light bulbs on the least busy afternoon.

The agent analyzes the calendar and adds the block to the least busy afternoon. Before the request, the user should record the calendar's current state to confirm the change.

Level 3 Checkpoint

Connect one tool used every day and assign the agent one task inside that tool.

Level 4: Train the Agent

AI models do not retain information between messages. Each message is preceded by a fresh briefing that includes the agent's identity, what it knows about the user, its notes, and the conversation so far. Corrections in one session are usually lost in the next.

Hermes stores corrections persistently across all sessions and devices.

Step 1: Correct a Response

Example request: Write a short post about studio lighting for complete beginners.

The response contains headings and bullets and is longer than preferred. Give a correction:

Too wordy. I prefer short, punchy sentences. Three paragraphs max. No headings or bullets.

The agent confirms that the preference was saved to memory. The next draft follows the stated format.

To test persistence within the same session, request a second post: Now, write one about microphones for video production. The output matches the first corrected style.

Step 2: Test Persistence in a New Session

1. Start a new session by clicking New session in the left menu, or press Command + N (Control + N on Windows).
2. Request an unrelated post: Write me a short post about roasting coffee.
3. The output follows the same structure, even though the session has no prior conversation.

The preference is stored in memory and applies to every session on every device.

Step 3: Review Memory Files

Settings:

1. Open settings with Command + , or the gear icon.
2. Select Memory and context in the left menu. Two main settings appear: Persistent memory and User profile.

File access in the web dashboard:

1. Open the Files section in the left navigation.
2. Open the memories folder.
3. Two Markdown files appear:
   • MEMORY.md holds the agent's working notes.
   • USER.md holds the user profile.
4. Click Download to save either file. The files open in any Markdown-compatible application.

Memory contents: Examples include the Google Workspace account's email address and the tools available in that workspace. Each time the agent starts, it reads its memory to determine what resources it can use, then reads the incoming message to determine the task.

Structure:

The agent writes and updates both files. The user can read them and request specific changes.

Both files are short by design, usually a few hundred words each, because each word is sent with every message. Additional information is stored in two other places: a searchable history of all conversations and a set of saved files and skills.

Step 4: Create an Index

When memory files grow, an index file provides a table of contents to all stored notes.

1. Send a request: From now on, keep an index file of your notes. A table of contents that points to where each topic lives. And put a one line pointer to it in your memory.
2. After processing, the agent creates the index file and records the pointer in memory.
3. To view the index, request it in chat, or locate it in the web dashboard:
   1. Open the folder opt/data.
   2. Open NOTES_INDEX.md.

The short memory contains only a pointer to the index. The agent reads other files when a task requires them. The index updates automatically over time.

Step 5: Save a Procedure as a Skill

Memory stores facts and preferences. Skills store procedures. A skill is a written set of steps for one task. The agent reads a skill only when that task arises.

Create a skill from a conversation:

1. Open the earlier post-writing chat.
2. Enter: Save that editing workflow as a skill called concise-writer.
3. The agent saves the workflow as the concise-writer skill and reports the file location.

View skills:

1. Select Capabilities, then the skills list.
2. Skills the user requested or the agent created itself carry the Learned tag.
3. Skills without the tag are built into Hermes.
4. Click a skill to view its text. Skills are plain text files.

Adaptive behavior: Skills are not static. In one example, the agent was asked to check a supplier's price page weekly. After the supplier redesigned its website, the old steps no longer found the prices. The agent determined a new path, rewrote its steps, and succeeded on the next run. The agent also writes new skills when it solves a problem. A background process reviews the library and archives skills that are no longer used.

Step 6: Convert a Document into a Skill

1. Drag a document into Hermes. The example uses an SEO starter guide in PDF format. A webpage link also works.
2. Enter /learn and add a brief request, such as: Turn this PDF into a reusable skill.
3. The agent creates and verifies the skill. In the example, it was named google-seo-starter.

Examples of documents that can be converted: a company style guide or a course syllabus.

Step 7: Install Existing Skills

Pre-built skills are available for installation.

1. In Capabilities, open the Skills page. Scroll to the list of skills available to install.
2. Click Install on any skill of interest.
3. For a larger library, select Browse the full hub at the bottom of the Skills page. This opens the Skills Hub, a library of skills written by other users.

Caution: A skill consists of instructions the agent follows. Users trust the author of each skill. Hermes scans Skills Hub entries for hidden instructions and data theft, and it blocks dangerous skills. The scan does not catch every risk. Read a skill before relying on it, or ask the agent to review it. Use official, trusted sources during the learning period.

Step 8: Set a Personality

Temporary personality (per session):

1. In any chat, enter /personality.
2. Select a preset from the list. Options include helpful, concise, creative, teacher, catgirl, and pirate.
3. Confirm the selection. The agent replies in the selected style for that session.

To restore the default, send /personality again and select none.

Permanent personality: The file SOUL.md defines who the agent is and how it behaves. Memory stores what the agent knows about the user. The soul defines the character it brings to the work. The file should remain short.

To change it, make a request: Update your SOUL.md so you always answer in a calm, direct tone. And never end a reply with a question unless I asked for options. After the update, the agent avoids follow-up questions at the end of replies.

Level 4 Checkpoint

Correct a recurring error. Tell the agent one fact about the user that it should always retain. Or convert a document the user explains frequently into a skill. Each step takes about one minute, and the user does not need to repeat the instruction afterward.

Level 5: Automate a Recurring Task

Hermes includes 16 prebuilt automations, including a morning briefing, a news digest, and a weekly review. Each of these sends the user something to read. The automation in this level performs a task instead.

Example: Receipt Filing

The example task is filing receipts. Each purchase confirmation and invoice in the user's email must be recorded in a spreadsheet, a task that takes about 20 minutes each week.

The automation runs on the agent's own Google account, created in Level 3. It never accesses the user's personal inbox.

Prompt used:

Set up a receipt filing clerk for me. First, set up a Google Sheet called Expenses with a Receipts tab if I don't have one yet. Then every day, at 7:00, search my Gmail for purchase receipts, order confirmations, and invoices from the last two days that aren't already labelled Files. For each one, pull out the vendor, the date, the total, and the currency, and append a row to my Google Sheet Expenses on the Receipts tab with columns for date, vendor, amount, currency, category, and the email ID. Then add the Files label to that email so it never gets counted twice. This is a pre-approved recurring job. So append the rows and apply the label without asking me each time. Never send, reply to, forward, delete, or archive anything. Read and label only. When you're done, message me one line on Telegram with how many you filed and the running total for the month. Send me a link to the spreadsheet that I can open from any device. Create the job paused so that I can look at it before it goes live.

Prompt components:

1. Permitted actions are stated. The agent may read and label only. All permitted actions are additive. The worst outcome is an incorrect row, which can be deleted.
2. The label serves as a record. Each run starts with no memory of the previous run. The Files label identifies emails already processed, which prevents duplicate entries.
3. The job is created paused. The agent builds the job but does not activate it, so the user can review it first.

Execution:

• The agent completed setup, testing, and a first run in about 10.5 minutes.
• The spreadsheet listed all receipts.
• The inbox showed all processed emails with the Files label.

Activation:

1. Open Scheduled jobs in the left menu. A paused job is marked with an orange indicator and the word "paused."
2. The job contains a readable prompt. The agent generated a deterministic script, so the job does not call the language model on each run. This reduces token costs. Each run consumes credits only for the initial trigger.
3. Click Resume to activate the job.
4. Optionally, click Trigger now to run it manually. In the recording, the manual trigger returned an error stating that the job was already running or claimed by another scheduler. The job ran anyway.

After activation, the job runs daily at 7:00 p.m. and processes the previous two days of emails.

Other scheduled-job candidates: Any recurring task that follows a rule and produces a file, folder, or calendar entry. Examples include a tracker updated from confirmation emails, preparation blocks added before meetings, or a weekly folder containing the week's notes.

Example: Flight Price Monitoring

A scheduled job can also monitor a condition and notify the user only when it changes.

Prompt used:

Watch Barcelona to London flight price trends for me. Build yourself a reusable skill for the check. Keep track of what you've already shown me. Schedule it every few hours. And message me on Telegram only when there's something new, written with your concise-writer skill.

Results:

• Setup took about 30 to 45 minutes. The example task was less specific than the receipt prompt, so the agent researched the method on the live internet before building the job.
• The agent created a skill for the check, a schedule, and a memory of prior results.
• The job is a deterministic script that runs every six hours and fetches current prices.
• Telegram notifications arrive only when new results differ from earlier results.

The agent in the example runs more than 20 scheduled jobs for its owner. They include a weekly bookkeeping check on Mondays, a birthday tracker that begins gift research early enough to order in time, a trip planner that runs twice per day and returns the best quotes, and a negotiation follow-up job.

Level 5 Checkpoint

Select one weekly task that ends in a spreadsheet, folder, or calendar. Instruct the agent to perform it. Create the job paused, review its contents, and then activate it.

Level 6: Build a Team of Agents

Subagents for a Single Task

A single request can divide a job among temporary helpers, called subagents. The helpers run at the same time and return one combined result.

Example request:

Plan a three-day trip to Lisbon, Portugal from London, leaving November 5th. One subagent on things to do. One on best places to eat for authentic local food. And one on the best flights itinerary. Then give me a single combined plan.

The main agent delegates the three tasks to three subagents. Each subagent runs on DeepSeek V4 Flash. Each subagent can also run on a different model if specified. The results are combined into one itinerary. Subagents exist only for the duration of the job.

Permanent Teammates: Profiles

A permanent teammate is called a profile. Each profile is a separate agent on the same server, with its own memory, skills, personality, and model. The desktop app presents profiles as bots with a name, title, job description, and avatar.

Job descriptions: Ask the main agent to write them, because it understands the board's requirements. Example request:

I'm about to create three teammate profiles on this server for a project board, a researcher, a writer, and a reviewer. Write a one-sentence job description for each one, worded so the board can tell from the description alone which of them a task belongs to. Give me three lines that I can copy and paste.

The board uses these descriptions to assign tasks.

Creating a bot:

1. In the upper left, switch from Sessions to Bots. If the tab is not visible, search the app or ask Hermes where it is.
2. Click the plus button, then New bot.
3. Assign an avatar. Options include a generated face from a text description, an uploaded picture, or a selection of pets.
4. Enter a name and title, such as researcher. Paste the job description.
5. Click Advanced to set additional options. If unchanged, the new bot copies the default profile.
6. To set a specific model, select a provider and model. The example sets the researcher to Opus 5 from Nous Portal, because it performs in-depth, long-running research.
7. Click Create bot.
8. Repeat the process for the writer and the reviewer, selecting models suited to each role.

The example used a top model for the researcher, because judgment is its primary function, and lower-cost models for the other two. Different models were used across the team to reduce bias from a single model type.

Test: Send Hello. to the reviewer. It responds that nothing is queued for review and is ready when needed. The agent already knows its role.

The Kanban Board

Hermes includes a project board named Kanban. It works like Trello: columns represent stages, and cards move from left to right. Unlike Trello, the cards do not wait for user action. The board divides the job into steps, assigns each step to a teammate, and moves the cards through the columns.

Enable the board:

1. In the desktop app, select Capabilities, then the Plugins tab.
2. Toggle Kanban on. It appears in the left column.

In the web dashboard, Kanban appears at the bottom of the navigation list, under Plugins.

Columns:

The board sets three of these columns automatically. Cards cannot be dragged into them.

Dependencies: A card can be blocked by another card. A dependent card cannot start until its prerequisite card reaches Done.

Notifications for Board Tasks

Tasks run in the background, so a stuck card can go unnoticed. Two methods prevent this:

1. Create cards from Telegram. A card created with the Telegram kanban command subscribes the user automatically. The user receives a message on completion or when the card becomes stuck.
2. State notification instructions in the job. Specify in plain language whom the agent should contact if it becomes stuck, as with any human team member.

Telegram command procedure:

1. In Telegram, type /kanban. The command does not appear in the suggestion list, because it is less common. Continue typing.
2. Ask Hermes for an example if needed: Show me an example of how to write a create command for the Kanban board.
3. The title is the only required argument. The syntax is /kanban create "Title" -- Assignee.
4. Example: /kanban create "TEST TASK" -- Default. The Default assignee sends the task to the default agent.
5. Telegram may merge the two dashes. Type -, a space, and -, then delete the space to produce two dashes.
6. Submit the command. The task is created and the user is subscribed to it.
7. Refresh the desktop app with Command + R (Control + R on Windows). The task appears in the Running column.
8. Upon completion, Telegram sends a notification, and the task moves to Done.

Create a Board Task in the Desktop App

Complex tasks are better assigned as a single description, with the agent determining the breakdown. Example:

Task title: Lisbon Trip Plan

Task description:

Plan a five-day trip to Lisbon for two people in early spring on a mid-range budget. Research flights from London, three neighbourhoods to stay in, and three day trips, all with real prices and a direct booking link for anything I'd have to buy. Build a budget spreadsheet in my Drive Outbox folder with a row for every item. What it is, the price, the booking link, a booked column, and an actualized column for what I really paid. Write a day-by-day itinerary next to it. Put placeholder events on my calendar for the travel days and anything time specific. Check everything for errors and missing prices before you show me. Message me on Telegram when it's done. And message me there if you get stuck and need an answer from me.

Execution:

1. Click New task, enter the title and description, then click Create task.
2. The task enters the Triage column while the orchestrator agent divides it into subtasks. Complex tasks take longer in Triage.
3. Subtasks are then assigned. In the example, three subtasks ran at once: flight research, neighbourhood research, and day trip research.
4. Subtasks with dependencies are marked as blocked by their prerequisites. For example, placeholder calendar events depended on the flight research.
5. Each card keeps its own history and comments. Activity logs display the work performed, including commands running in real time.
6. Comments can be added to any card. The agent and the orchestrator also add comments as cards move between columns.

In the example, the research tasks were assigned to the researcher. The budget task kept the default assignee. The review task was assigned to the reviewer. Each assignee was chosen based on its specialization.

Tasks That Wait for External Events

A single task cannot wait for an external event, such as a hotel email, because each task runs and then ends. A scheduled job resolves this:

1. Create a scheduled job that checks the board. In the example, the check runs every 30 minutes.
2. The job reads blocked cards. When the awaited event occurs, it moves each card back to Ready, and work resumes.
3. If a card requires user input, the job sends a Telegram message instead of leaving the card blocked without notice.

Example prompt for the booking-confirmation job:

Every morning, check my email for booking confirmations for the Lisbon trip. For each one, fill in the actualized column in the trip spreadsheet with what I actually paid, tick booked, delete the matching placeholder event from my calendar, and create the real one with the confirmation number in the description. Add guest@mattexmatt.com as a guest on any event they're joining. Label the confirmation email Trip so it isn't counted twice. Message me one line on Telegram with what changed. Create it paused so I can read it first.

Caution: Adding a guest to a calendar event sends that person a real invitation. Names should be added only when the user intends to invite them. The job should be checked while paused.

A board performs the work. A scheduled job keeps the work current and signals when user input is needed. The same structure applies to a weekly report, a batch of product descriptions, a content calendar, or a research and outreach list.

Maintenance and Administration

Diagnostics

Use the three checks from Level 2 for any malfunction:

1. Credits
2. Gateway
3. Logs

Updates and Backups

Dashboard system page:

1. In the web dashboard, scroll the left navigation and select System.
2. The page displays an overview of the setup and its current state.
3. The gateway section shows whether the gateway runs. The gateway can be stopped, restarted, or started from this section.
4. The Backup section creates backups with a button. Create a backup before any major configuration change.
5. The same section restores from a backup. A downloaded backup file can be uploaded to restore the saved state.

Server update:

1. In Hostinger Docker Manager, click the three dots to the right of the application.
2. Click Update. A notification confirms that the application is updating.
3. This updates the Hermes version on the VPS, known as the backend.

Desktop app update:

1. Open settings.
2. Select About in the left menu.
3. Check for updates. If available, click Update now.

The desktop application and the server software update separately.

Achievements

The web dashboard includes an Achievements page under Kanban in the Plugins section. Hermes reviews the agent's history and awards badges.

• 60 badges in total, across categories including agent autonomy, debugging chaos, and tool mastery.
• Tiers range from copper to Olympian.
• Some badges remain hidden until triggered.

At the time of recording, the example account had unlocked 7 of 60 badges, had 53 undiscovered, and had reached the gold tier.

Model Selection and Cost Management

Choosing a Model

Models function as workers hired for specific jobs. High-capability models cost more. Low-cost models perform routine tasks. Each model can also operate at different effort levels.

Model catalog: In Nous Portal, under Resources, select Models. The page shows:

• Active promotions, with standard and current prices.
• The full list of models, with filters and portal prices.

Example prices (per million tokens):

Claude Fable 5.1 is a high-capability model. Free models are also listed, but performance varies. Experimentation determines which models suit a given task.

Matching model to task:

• Routine, skill-based tasks, such as checking a feed or filing receipts, can run on a low-cost model. Avoid the cheapest and free options for these tasks, because a model that cannot reason produces confident errors, such as filing the wrong item or discarding a needed item.
• Judgment-based tasks, such as planning, analysis, or any task where error is costly, require a top model. Hermes performs noticeably better on flagship models, particularly those that handle large volumes of information.
• If results are weak, change the model or increase the effort level before concluding that the tool is inadequate.

Switching Models

In the desktop app:

1. Click the model selector to the right of the message box. It lists the models available through the connected provider.
2. Select a model before sending a message.
3. To change effort, click the dropdown next to the model selector. Effort levels include medium, ultra, and max. Medium suits most tasks. Ultra increases cost substantially on high-cost models.

In Telegram:

1. Type /model to open the model configuration.
2. Toggle to view available models.
3. Type /reasoning to select the reasoning level. Telegram labels effort as reasoning.

Cost considerations:

• Switching models in a long conversation causes the new model to reread the full conversation, which costs more. For large tasks, start a new thread on the new model.
• Each scheduled job records the model it was created with. If the default model changes, an unpinned job refuses to run rather than switching to a more expensive model. This behavior is intentional. To resolve it, pin the job to a model on the Cron page:
  1. Open the scheduled job list in the desktop app.
  2. Click the three dots next to the job, then Edit cron.
  3. For jobs that call a language model, select a model from the model selector.
• Jobs built as deterministic scripts do not call a language model, so no model selection is required for them.

Fallback Chains

Models sometimes stop responding, and plans reach usage limits. These failures often occur during scheduled jobs when no user is present. A fallback chain provides an ordered list of models the agent uses when the primary model does not respond.

Example structure:

1. Flagship model
2. Model from another provider
3. Low-cost, reliable model, which can be a free option as a last resort

The agent switches models during a conversation without losing context. Scheduled jobs use the same list.

Setup procedure:

1. Start a new session.
2. Enter a request, substituting the user's own models: Set up a fallback chain. Try flagship model first. Then a model from another provider. Then a cheap reliable model.
3. Select model names from the Nous Portal model list. The example used:
   • Primary: DeepSeek V4 Pro
   • Second provider: Claude Sonnet 5
   • Last resort: Upstage Solar Pro 4
4. The agent confirms that the chain is configured.

Mixture of Agents

The /moa (mixture of agents) command sends a request to several models at once and combines their answers into one reply. It multiplies cost, since several answers are generated. It suits decisions that benefit from multiple independent opinions.

Configuration:

1. If the /moa command returns a provider error, open the switch provider prompt.
2. In the model settings, scroll to the Mixture of agents section.
3. Select reference models. The example used:
   • Reference 1: Opus 5 (Anthropic)
   • Reference 2: ChatGPT 5.6 Luna
   • Reference 3: DeepSeek V4 Pro
4. Select the aggregator, the model that combines the reference answers. The example used DeepSeek V4.1 Flash.
5. The fallback models can also be managed manually from the bottom of the same page.

Example test: The request asked whether the chosen models were the best options for a fallback chain consisting of a flagship model, a reliable model, and a cheap or free model.

Result: The example session took about 26 minutes. The output recommended this fallback order:

1. Opus 5
2. DeepSeek V4 Pro
3. DeepSeek V4 Flash
4. Solar Pro, as the last resort

The recommendation applies to the example setup only. Models change frequently. Users should test models against their own tasks.

Cost Monitoring

Usage page: In Nous Portal, under Account, select Usage. The page shows:

• Total credit spend
• Total input tokens
• Total output tokens
• Cache reads and cache writes
• Total number of requests

Usage can be viewed by time period, by day, by model, and by chart style.

Token definitions:

• Input tokens: text sent to the model, including prompts.
• Output tokens: text generated by the model.

Both consume credits. Each message resends the agent's full briefing, including its instructions, memory, tool list, and conversation history. This repeated reading is called context, and it accounts for most of the bill. The portal recognizes unchanged portions of context and charges a fraction of the standard rate to read them again. This process is called caching and requires no setup.

Cost observations from the recording:

• The first day of testing cost slightly more than eight cents.
• The multi-model mixture-of-agents task cost $1.30 for calls to Opus 5. Earlier usage, which relied on DeepSeek V4 Flash, cost far less.
• Setting Anthropic's Claude Fable 5.1 as the default for every task would produce high costs. Such models suit tasks that demand them.
• Most daily work benefits from a model that balances cost and performance.

Server cost: The Hostinger server bill is fixed. The model bill varies with usage and requires monitoring.

Summary of Levels

Start with small tasks. Let the agent learn working patterns, and expand its responsibilities as trust develops.