Hermes Agent glossary
75 terms you will meet in the docs and the community, defined in plain English.
.
- .env file
The file in your Hermes home (
~/.hermes/.env) that stores API keys and tokens such as TELEGRAM_BOT_TOKEN. It holds credentials, not memory.FAQ & Troubleshooting (opens in a new tab) Persistent Memory (opens in a new tab)
/
- /goal
A standing goal Hermes keeps working toward across turns. After each turn a judge model decides whether it's done, and if not Hermes continues, up to a budget of 20 turns by default (
goals.max_turns).- /steer
A slash command that hands the running agent a note after its next tool call, changing direction without interrupting the task or starting a new turn.
A
- ACP (Agent Client Protocol)
A JSON-RPC-over-stdio protocol spoken by editors such as VS Code, Zed and JetBrains.
hermes acpruns Hermes as an ACP server so it can work inside those editors.ACP Host Integration (opens in a new tab) Programmatic Integration (opens in a new tab)
- AGENTS.md
The main project context file describing structure, conventions and instructions. Inside a git repo,
AGENTS.mdfiles from the root down to the working directory are merged.- API server
A mode that exposes Hermes as an OpenAI-compatible HTTP endpoint so other chat frontends can use it as a backend.
- approvals.deny
A user-editable list of glob patterns for terminal commands that are always blocked, even in YOLO mode or with approvals turned off. It's the configurable counterpart to the hardline blocklist.
- Auxiliary models
Models used for side tasks such as vision, context compression, session titles and approval checks. They default to your main chat model and can be overridden per task under
auxiliaryinconfig.yamlor throughhermes model.
B
- Baileys bridge
Hermes' built-in WhatsApp connection, which emulates a WhatsApp Web session paired by QR code rather than using the official Business API.
- Bot Chat
A profile's canonical chat session with its own bot. The
bot-chatcron delivery target posts output into it as an incoming message so the bot acts on it, costing one full agent turn per delivery.
C
- Checkpoints / rollback
An opt-in feature (
hermes chat --checkpoints) that snapshots your project before destructive operations so you can restore it with/rollback.- command_allowlist
The
config.yamllist of command patterns you've permanently approved, so Hermes runs them without asking.- config.yaml
The main settings file in your Hermes home. It holds the model, terminal backend, memory, approvals, MCP servers, display options and more, and you can edit it directly or use
hermes config set.Configuration (opens in a new tab) FAQ & Troubleshooting (opens in a new tab)
- Context compression
Automatic summarising of older conversation history once a session passes a share of the model's context window. You can also trigger it yourself with
/compress.Configuration (opens in a new tab) FAQ & Troubleshooting (opens in a new tab)
- Context file injection scanning
A check that scans
AGENTS.md,.cursorrulesand similar files for prompt-injection patterns before they enter the system prompt. A flagged project file is not loaded and a warning is shown instead.- Context files
Project instruction files Hermes loads automatically, such as .
hermes.md,AGENTS.md,CLAUDE.mdor .cursorrules. Only the first matching type is used per session.- Context references
Inline @ syntax such as
@file:,@folder:,@diff,@staged,@git:and@url:that pulls file contents, git changes or web pages straight into a message.- Credential pool
Several API keys for the same provider that Hermes rotates between when one hits a rate limit or quota.
- Cron job
A scheduled task, one-off or recurring, that runs in a fresh agent session and delivers results to a chat or file. It's managed through
/cron,hermes cronor plain-language requests.- cronjob_manage
The single tool Hermes uses to create, pause, edit, trigger and remove scheduled jobs when you ask in plain language.
- Curator
A background maintenance pass that tracks how agent-created skills are used, moves unused ones from active to stale to archived, and never auto-deletes them.
- Custom endpoint
A provider option (provider: custom) for any OpenAI-compatible server, commonly used for local models through Ollama, vLLM or llama.cpp.
D
- Dangerous command approval
The safety check that pauses risky shell commands for your approval. approvals.mode sets it to smart, manual or off.
- delegate_task / subagent
The tool that spawns child agents with isolated context for side tasks or parallel work. Only each child's final summary returns to the main chat.
- Delivery target
The
deliversetting that decides where a cron job's output goes:origin,local, a platform's home channel, a specific chat or topic, a comma-separated list, orall.- DM pairing
A way to authorise new chat users: an unknown user gets an 8-character code, which the owner approves with
hermes pairing approve.
E
- execute_code
A tool that lets the agent write a Python script that calls Hermes tools programmatically, collapsing many steps into one turn.
- External memory provider
A plugin-based memory service, such as Honcho, Mem0, Hindsight, Supermemory, RetainDB or ByteRover, that adds deeper recall alongside the built-in memory files. Set it up with
hermes memory setup.
F
- Fallback providers
Backup provider and model pairs Hermes switches to automatically when your main model fails.
- Frozen snapshot
How memory is loaded: it's read once into the system prompt when a session starts and doesn't change mid-session. That preserves the prompt cache but means new memories appear in the next session.
H
- Hardline blocklist
An always-on list of catastrophic commands, such as filesystem wipes and fork bombs, that Hermes refuses even in YOLO mode.
- Hermes Agent
An open-source (MIT) AI agent from Nous Research that runs in your terminal, a desktop app or chat apps, uses tools for you, and learns over time through memory and skills.
- Hermes Desktop
The native desktop app for macOS, Windows and Linux. It drives the same agent, config and sessions as the CLI.
- Hermes home (HERMES_HOME)
The data folder where Hermes keeps config, keys, memory, skills and sessions. It's
~/.hermeson Linux and macOS, and the HERMES_HOME environment variable points Hermes at a different one.Installation (opens in a new tab) FAQ & Troubleshooting (opens in a new tab)
- Home channel
The default chat on a platform where cron results and notifications land. Set it by sending
/sethomein that chat or with variables such asTELEGRAM_HOME_CHANNEL.Telegram (opens in a new tab) Scheduled Tasks (Cron) (opens in a new tab)
- Honcho
An external memory provider from Plastic Labs that builds a model of the user across sessions and profiles. It's installed with
hermes plugins installhoncho and runs on Honcho Cloud or self-hosted.Memory Providers (opens in a new tab) Hermes Agent README (GitHub) (opens in a new tab)
K
- Kanban board
A durable task board in
~/.hermes/kanban.db, shared across your profiles, where named agents pick up, hand off and complete tasks. Agents work it throughkanban_*tools, while you usehermes kanban,/kanbanor the dashboard.- Kanban dispatcher
The loop that promotes ready Kanban tasks, claims them and spawns the assigned profile as a worker, reclaiming crashed or stale workers along the way. It runs inside the gateway by default and ticks every 60 seconds.
M
- MCP (Model Context Protocol)
An open protocol for external tool servers. Hermes connects to MCP servers listed under mcp_servers in
config.yamland uses their tools like built-in ones.- MCP server mode
Running
hermes mcp serveturns Hermes itself into a stdio MCP server, so clients such as Claude Code or Cursor can list conversations, read history and send messages across its connected platforms.- Memory tool
The tool the agent uses to add, patch or remove memory entries. A memory only persists if the model actually calls it.
- MEMORY.md
The agent's own notes file in
~/.hermes/memories/, covering environment facts, conventions and lessons learned. It's capped at 2,200 characters by default and loaded at the start of every session.- Messaging gateway
The single background process that connects Hermes to Telegram, Discord, Slack, WhatsApp and other platforms, and also runs cron jobs.
- Mixture of Agents (MoA)
A virtual
moaprovider whose named presets appear as selectable models. Reference models analyse the turn first, then an aggregator model writes the reply and makes tool calls, and the aggregator's provider is billed for the run.
N
- No-agent mode
A cron option that runs a script on a schedule and delivers its output verbatim, with no LLM involved.
- Nous Portal
Nous Research's subscription service, giving one login to 300+ models plus the Tool Gateway. It's set up with
hermes setup --portal.- Nous Research
The AI research company that builds and maintains Hermes Agent and runs the Nous Portal subscription service.
Hermes Agent README (GitHub) (opens in a new tab) Nous Portal (opens in a new tab)
P
- Plugin
A Python package in
~/.hermes/plugins/with aplugin.yamlmanifest that adds custom tools, hooks or integrations without changing Hermes' core code.- Profile
A separate Hermes home under
~/.hermes/profiles/with its own config, keys, memory, sessions and skills, letting you run several independent agents.- Progressive disclosure
Hermes' token-saving way of loading skills: the agent first sees only names and descriptions, and opens a skill's full content only when needed.
- Prompt caching
Reuse of an unchanged conversation prefix so repeated input is billed more cheaply. Hermes enables it automatically for Claude models via Anthropic or OpenRouter and avoids rebuilding the system prompt mid-conversation, while a
/modelswitch resets the cache.Context Compression and Caching (opens in a new tab) FAQ & Troubleshooting (opens in a new tab)
S
- Secret sources
External secret managers (Bitwarden Secrets Manager, 1Password, or a command helper for other vaults) that inject API keys at startup, so only a bootstrap token sits in
.env. They're configured undersecretsinconfig.yaml.- Session
One stored conversation with its full message history, from the CLI or any chat platform. Sessions are kept in
~/.hermes/state.db.- session_search
The agent's tool for searching and scrolling through past conversations stored in
state.db. It costs no extra LLM calls.Persistent Memory (opens in a new tab) Sessions (opens in a new tab)
- Skill
A folder with a
SKILL.mdfile containing on-demand instructions for a task, compatible with the agentskills.io standard. Skills live in~/.hermes/skills/and work as slash commands.- skill_manage
The tool the agent uses to create, update and delete its own skills. The docs call this its procedural memory.
- SKILL.md
The required main file of a skill, holding its name, description and step-by-step instructions. It can be accompanied by references, templates, scripts and assets.
- Skills Hub
The browsable catalog of installable skills from official,
skills.sh, GitHub and other sources, managed withhermes skills browse, search and install.- Smart approval mode
The default approval mode, where a helper model auto-approves low-risk commands, denies clearly dangerous ones and asks you about uncertain cases.
- SOUL.md
A global persona file in your Hermes home that defines the agent's identity and tone. It's the first thing in the system prompt.
Personality & SOUL.md (opens in a new tab) Context Files (opens in a new tab)
- state.db
The SQLite database in your Hermes home that stores session metadata and message history with full-text search.
Sessions (opens in a new tab) Persistent Memory (opens in a new tab)
T
- Telegram privacy mode
A Telegram bot setting, on by default, that hides ordinary group messages from bots. Turn it off in @BotFather and re-add the bot, or make the bot a group admin, so Hermes can see group chat.
- Terminal backend
Where the agent's shell commands run: local, docker, ssh, singularity, modal, daytona or vercel_sandbox. It's set with terminal.backend.
- Tool
A single action the agent can take, such as running a terminal command, reading a file, searching the web or saving memory.
- Tool Gateway
Part of a paid Nous Portal subscription that handles web search, image generation, text-to-speech and cloud browsing without separate vendor accounts.
- Tool Search
An opt-in feature that loads tool schemas only when relevant, saving context when many MCP or plugin tools are attached.
- Toolset
A named bundle of related tools, for example file or terminal, that you switch on or off together with
hermes tools.Toolsets Reference (opens in a new tab) Tools & Toolsets (opens in a new tab)
- TUI
Hermes' modern terminal user interface, launched with
hermes --tui, which the docs recommend for interactive use.
U
- User allowlist
A list of platform user IDs, such as TELEGRAM_ALLOWED_USERS, that are allowed to talk to your bot.
- USER.md
The user-profile memory file in
~/.hermes/memories/for your preferences and communication style. It's capped at 1,375 characters by default.
W
- wakeAgent gate
A way for a cron job's pre-check script to skip the agent: if the script's last stdout line is
{"wakeAgent": false}, that run spends no LLM tokens. Leaving it out means the agent runs as usual.- Webhook adapter
The gateway component that runs an HTTP server (port 8644 by default), validates HMAC signatures on POSTs from services like GitHub or Stripe and turns them into agent runs. Routes come from
config.yamlorhermes webhook subscribe.- Website blocklist
The
security.website_blocklistsetting lists domains the agent may not reach. It's enforced acrossweb_search,web_extract,browser_navigateand every other URL-capable tool.- Write approval
An optional setting (memory.write_approval, and skills.write_approval for skills) that holds the agent's writes for your review. You manage them with commands like
/memorypending and/memoryapprove.Persistent Memory (opens in a new tab) Skills System (opens in a new tab)
Y
- YOLO mode
A per-session switch (--yolo or
/yolo) that skips approval prompts. The hardline blocklist still applies.
No terms match that filter.