Wiki · Start here
Glossary
Twenty-six words used around the agent — no jargon and no acronyms.
📖
The words people use around the agent — in plain language, no jargon.
When you meet one somewhere else and are unsure, come back here. What this page says holds on the other
wiki pages too.
The basics
| Term | What it means |
|---|---|
| Agent | A program you brief in plain language that then carries the work out itself — runs commands, reads and changes files, searches the web. It does not just answer, it acts. |
| Model (language model) | The "brain" that understands text and decides what to do. It usually lives in someone else's service and you pay for its usage. Models differ in ability, speed and price. |
| Model provider | The company you rent a model from (you pay per usage). You choose by price and capability. See choosing a model. |
| Key (API key) | The secret the agent identifies itself with at the model provider — like a password. It belongs in the agent's settings, never in texts, skills or messages to people. |
| Token | A chunk of text that usage is counted in (roughly three quarters of a word). You pay for the tokens the agent sends and receives. |
| Context | How much text the model holds at once — the live conversation including instructions and replies. When it grows past what the model can hold, the conversation must be trimmed or restarted. |
Talking to the agent
| Term | What it means |
|---|---|
| Prompt | What you want from the agent — an ordinary sentence, no special syntax. A specific prompt ("what, from where, to whom, how long") turns out better than a vague one. |
| Slash command | A short instruction starting with a slash that controls the agent directly — for example /new (fresh conversation), /compress (trim the conversation), /learn (learn a procedure). |
| Conversation (session) | One continuous exchange with the agent. It keeps its own running memory, and when it gets long the agent trims it so the important parts still fit. |
| Evidence | What a claim rests on — command output, file contents, a measured number. Our rule: a claim without evidence is not accepted. |
What the agent can do (and what it is made of)
| Term | What it means |
|---|---|
| Tool | A core ability of the agent: run a command, read a file, search the web, send a message. Tools are available from the start. |
| Skill | A written procedure for a particular job ("how we back up the database here"). The agent reads it when it needs it. See Skills. |
| Plugin | A code extension of the agent itself (a new feature, written in Python). Day-to-day use rarely needs one — skills are usually enough. |
| Connection (MCP) | Linking the agent to a service's ready-made interface (GitHub, an automation tool, a browser). A connection is third-party code — take it from trusted sources. See MCP tools. |
| Gateway | The part of the agent that carries the link to Telegram, Discord and other platforms. With it down, the agent does not answer messages from your phone — and scheduled jobs do not fire either. |
| Scheduled job (cron) | A task the agent performs by itself on a schedule (a morning brief, a nightly check). See Scheduled tasks. |
Where it all lives
| Term | What it means |
|---|---|
| Server (VPS) | A computer that is always on and connected to the internet — the agent's home. Its advantage over a home computer is that it never shuts down. |
| Terminal | The window where commands are pasted (and confirmed with Enter). Nothing happens on its own — you paste every command. |
| The agent's home folder | ~/.hermes/ — conversations, memory, skills, settings and keys live there. Backing up means backing up this. |
| Repository (git) | A folder with a history of changes — every edit is recorded and can be rolled back. Used for websites and code, not for the agent's settings. |
Safety and operation
| Term | What it means |
|---|---|
| Approval | The confirmation the agent asks for before a dangerous command (deleting, overwriting). Nothing like that happens without your consent. |
| Allowlist | The list of accounts allowed to control the agent over Telegram or Discord. Anyone outside it is not obeyed. |
| Pairing | The first safe contact with a new person over a direct message — only then does the agent listen to them. |
| Sandbox | A separated environment work can run in, so it cannot reach the rest of the server. Where the agent may write can be narrowed as well. |
| Telemetry | Collection of usage data. Hermes Agent sends none — your conversations, memory and skills stay with you. |
| Backup | A copy of data you can restore from. For the agent you back up ~/.hermes/; for a website, the whole site. |