Introduction
Hermes groups its built-in tools into seven primary categories, plus integrations for external systems. The categories are not bureaucracy. They map directly to capability and risk: each one describes what the agent can do and what surface it touches. If you remember the seven buckets and the integrations slot, you can place any tool you encounter without looking it up.
Key Concepts
- Category: A high-level grouping of tools by capability area.
- Capability surface: The thing the tool touches (your disk, the network, a browser, a model).
- Integrations: A separate slot for tools that connect to external systems (Home Assistant, MCP, Spotify, Discord admin, ...). Not one of the seven primary categories, but always present in the registry.
- Cross-category tools: A few tools (like
browser_vision) appear in more than one bucket because they span surfaces. That is normal.
Real World Context
When a developer asks does Hermes have a tool for X, the fastest answer is: which category would it live in. Send a Slack message is Automation and Delivery. Read a PDF from the web is Web. Run a Python snippet is Agent Orchestration (specifically execute_code). The category narrows the search from seventy tools to three or four.
Deep Dive
The seven primary categories, with their main tools:
- Web:
web_search,web_extract. Search the web and pull markdown from individual pages or PDFs. - Terminal and Files:
terminal,process,read_file,write_file,patch,search_files. Run shell commands, manage background processes, and manipulate files on the local filesystem. - Browser:
browser_navigate,browser_snapshot,browser_click,browser_type,browser_vision, and friends. Drive a real browser when the page requires JavaScript, login, or interaction. - Media:
vision_analyze,image_generate,text_to_speech,video_analyze. Multimodal analysis and generation: read images, make images, speak text, watch videos. - Agent Orchestration:
todo,clarify,execute_code,delegate_task. Planning, clarifying with the user, executing Python that itself calls Hermes tools, and spawning subagents. - Memory and Recall:
memory,session_search. Persistent cross-session storage and search across past conversations. - Automation and Delivery:
cronjob,send_message. Scheduled tasks and outbound messaging on platforms like Telegram, Discord, Slack.
And alongside these seven, the registry includes Integrations: ha_* (Home Assistant), MCP server tools, Spotify, Discord admin, and so on. Anything that talks to a specific external system. Integrations are not one of the seven primary categories because their contents depend on what you have configured, but they always occupy this slot in the registry.
This lesson is the map. The risk gradient in Section 3.2 and the web and media specifics in Section 3.3 zoom in on the categories that have the most teeth.
Note: Orchestration tools (delegate_task, todo, clarify) and Automation tools (cronjob, send_message) are deep-dived in Course 12. We mention them here so you know they exist.
Common Pitfalls
- Treating categories as airtight:
browser_visionstraddles Browser and Media.web_extractcan also handle PDFs. The categories are guideposts, not walls. - Forgetting Integrations exists: When an MCP server is loaded, its tools land here. If the agent suddenly has a new tool, check whether an MCP was added.
Best Practices
- Think in categories first: When you imagine an agent capability, name the category before naming the tool. It scales as the registry grows.
- Pair categories with risk: Web and Media are mostly read-only and safe. Terminal and Files include the most dangerous tools. We will formalize this in Section 3.2.
Summary
- Hermes tools live in seven primary categories: Web, Terminal and Files, Browser, Media, Agent Orchestration, Memory and Recall, Automation and Delivery.
- Alongside the seven, the registry has an Integrations slot for external-system tools (Home Assistant, MCP, Spotify, ...).
- Categories describe both capability and surface area.
- Some tools span categories; the categories are guides, not strict containers.
- Knowing the categories shrinks the search space when reasoning about agent capabilities.
Code Examples
# Mapping a few common requests to their category and tool
- request: 'find papers on X'
category: Web
tool: web_search
- request: 'patch this file'
category: Terminal and Files
tool: patch
- request: 'log into the staging dashboard'
category: Browser
tool: browser_navigate (then browser_type, browser_click)
- request: 'describe this PNG'
category: Media
tool: vision_analyze
- request: 'send the report to #engineering at 9am daily'
category: Automation and Delivery
tool: cronjob (which then calls send_message)