🧩 Capabilities

A simple text file (markdown) that stores what matters about you — facts, preferences, context.

It's human-readable and lives on your machine; you can open and edit it whenever you want.

Plain text; transparent; portable.

An index that lets you search any message from any session by word or date.

You recover “what I asked on May 19” without relying on human memory.

Full recall; search by date; complements MEMORY.md.

Memory builds up in the background while you use it.

It takes ~2 weeks of use for the agent to become truly useful—patience pays off.

Compounding effect; the more you use it, the better it gets.

Connecting external sources (Obsidian, Pinecone, Claude Code memory, PC apps) expands what Hermes knows.

More sources = more accurate, proactive advice.

Expanded memory; knowledge integrations.

During idle periods, the agent processes what it has learned and proactively comes back with ideas.

Turns passive memory into initiative: it suggests things before you ask.

Idle processing; proactivity.

Without knowing what you do, it can't give precise advice. Memory is the foundation of everything.

Defines the philosophy: a useful agent is one that knows you.

Context = quality; everything system.

A file that turns a generic chatbot into an assistant with YOUR vibe.

Without a soul, the agent doesn't know who to be.

Identity; personality defined in text.

Defines the agent’s identity, mission, goals, and values.

Provides direction: the agent knows what to prioritize.

Mission; values; goals.

Sets the tone: it can say “bro,” “dude,” be formal, or be direct.

The conversation feels natural and works your way.

Tone; form of address; voice.

You can give the agent a character archetype.

A consistent character is more pleasant to use.

Archetype; consistency.

The soul is what makes your Hermes different from everyone else's.

Real customization, not the same chatbot as everyone else.

Uniqueness; customization.

There’s a template with identity, mission, goals, and business details to fill in.

You don’t start from a blank page.

Template; guided completion.

Hermes comes with ready-to-use integrations: text-to-speech, image generation, vision, web search.

You enable capabilities without writing any code.

TTS; image; vision; web.

Each integration only needs the service's key to work.

Simple setup, but the key needs to be stored securely.

One key per tool.

Everything you type in the chat is saved/indexed — a pasted key becomes a permanent risk.

A leaked key can cost you a lot. Golden rule.

Chat = logged; key stays out of chat.

Ask Hermes for a terminal command, replace the placeholder with your key, and paste it into the terminal.

The key stays outside the chat history, in the system environment.

env var; placeholder; terminal.

Granola integration: "what was the last meeting and 1 action item I promised?".

Shows an integration solving a real everyday problem.

Meetings; promised actions.

Tools like Granola MCP use OAuth: just provide the MCP name and authorize it in the browser.

No keys to manage; connect by logging in.

OAuth; authorization in the browser.

Since it runs on your machine, it creates and organizes files on the desktop.

Goes beyond answering: produces real artifacts.

File system access.

Uses native Chrome DevTools and visual screen reading to navigate.

It understands the page as you see it.

DevTools; screen vision.

Runs terminal commands to automate tasks.

Bash gives you real automation power over the system.

Commands; automation.

Moves the cursor and clicks for real — real agency, not just a screenshot.

Lets you interact with any app, even without an API.

Real movements; fused.

Forgot to do something on the way to the gym? Hermes runs it on your PC.

Shows the value of an agent that acts remotely.

Remote execution; convenience.

It's a powerful feature you enable in the tools — use it carefully.

The ability to act on your PC calls for the principle of least privilege (Track 3).

Responsibility; least privilege.

A pattern that packages a tool for the model to use — the manual wrapped around the API.

It's how Hermes connects to almost everything in a standardized way.

Protocol; standardization.

The API is the exposed wiring; the MCP is the package of instructions around it.

It's a "universal remote control" — an analogy that makes the concept clear.

Wiring vs. manual; universal control.

The model reads the manual and already knows all the buttons, without repeated checks.

Uses far fewer tokens and standardizes usage.

Token efficiency; less overhead.

Meeting MCP: accesses notes and action items via OAuth authorization.

Simple plug-and-play MCP example.

Meetings; OAuth.

Starts servers that connect to apps; you add tools (e.g., read email) and generate a token.

Opens thousands of apps — add the token via env, not in chat.

Token via env; selective tools.

Add “read email,” but hold off on “send/reply to email” until you trust the flow.

Least access prevents damage; never put tokens in chat.

Least privilege; secure token.

The more you use it and give feedback, the more Hermes iterates and improves.

Your usage trains the behavior — muscle memory.

Feedback; iteration.

Hermes has skills it gains through use, and you can build your own.

Skills are recorded workflows for how to do tasks.

Skills; agentskills.io.

You create personas (e.g., "the Alchemist") with a job, description, system prompt, and their own model.

Dedicated specialists for specific tasks.

Personas; model per persona.

Expensive models only for high-impact tasks; autopilot uses a low-cost model.

Saves money without sacrificing quality where it matters.

Cost-effectiveness; model allocation.

"Research the best country to live in": spins up a sub-agent, uses DeepSeek V4 + GPT in a loop, with Opus 4.7 reviewing.

Skill that delegates to different models depending on the role.

Delegation; model loop.

The command /stop stops the task in progress.

Control: you can stop whenever you need to.

Interruption; control.

Queues a prompt for the next turn without interrupting the current one.

You move on to the next step while it works.

Queue; next turn.

Runs tasks in parallel in the SAME chat — multiple concurrent workflows.

Does more things at once without opening multiple windows.

Parallelism; concurrency.

/canban opens a task board; /reset clear everything.

Organizing and starting fresh improves performance.

Kanban; reset.

Compresses memory/context: "summarize everything we’ve discussed" and continues with less.

Less context = better performance and lower cost.

Compression; summary.

Switch the model that’s thinking — the brain of Hermes.

Uses the best model for each task (see Track 1).

Model-agnostic; brain switching.

Each question uses ALL the context + the question; the larger it is, the worse the performance.

Ideal: one session, one window; reset and compact.

Context; reset; compress.