Tools, toolsets and the terminal backend
All of the agent's power comes from tools. A tool is a function the model can call: terminal, read_file, patch, web_search, browser_navigate, memory, cronjob and dozens more. Tools are grouped into toolsets so you can switch them on and off in bundles rather than one by one.
| Toolset | What is inside | Risk level |
|---|---|---|
| `terminal`, `file` | `terminal`, `process`, `read_file`, `patch` | High — makes real changes |
| `web`, `search` | `web_search`, `web_extract` | Medium — external content enters the context |
| `browser` | `browser_navigate`, `browser_snapshot`, `browser_vision` | Medium to high — it can log into sites |
| `memory`, `session_search` | Writing memory and searching past sessions | Low |
| `cronjob` | Creating scheduled tasks | High — it runs itself in the future |
| `delegation`, `code_execution` | `delegate_task`, `execute_code` | High |
| `safe` | A set of harmless tools only | Low |
The second important decision is the terminal backend: where the agent executes commands. It is one configuration line that completely changes the blast radius:
local— on your machine, with your user rightsssh— on a separate machine; the agent cannot modify its own codedocker— in a container, with hardened flags and resource limits- cloud sandboxes (
modal,daytona,vercel_sandbox) — isolation that scales
The default is local — so from day one the agent carries your user rights.
A subtlety worth noticing: the dangerous-command check runs on the local and ssh backends, and is skipped on container backends — because there the container is the boundary. Moving to docker is therefore not "more control" but "a different kind of control": it stops asking for approval, but the damage stays inside the container.
Practice. Run the same task on two backends: local and docker. The task: "create report.txt in the current directory and write the date into it". Then find where the file was created in each case. Done means: for docker you can show that the file is not on the host but in the sandbox directory.
📚 Sources and documentation
- Tools and toolsetsofficialhermes-agent.nousresearch.com
The full toolset list and the terminal backend settings.
- Security: backend comparisonofficialhermes-agent.nousresearch.com
Which backends run the dangerous-command check and which skip it.
- Configurationofficialhermes-agent.nousresearch.com