Optional deep dive · Configuration
Hermes has a lot of settings screens, and none of them were written for nurses. This handbook walks every one of them, in plain language, and tells you the safe default — so you can configure with confidence instead of clicking with crossed fingers.
How to use this handbook
For every setting, this handbook answers the same three questions: What is it, in plain language? What would a nurse call it? What should I set it to? Read it front to back once if you like, or jump straight to the screen you're staring at.
Before anything else
If you only keep one thing from this handbook, keep this table. It's the one-glance med sheet: the settings that matter, and the values that keep a beginner's Hermes useful enough to work and scoped enough to stay safe.
| Setting | Safe value | Why |
|---|---|---|
| Working Directory | your My-Nurse-AI-OS folder | Hermes works in one assigned unit, not the whole hospital. |
| Code Execution Mode | Project | Keeps code runs inside the current project. |
| Environment Passthrough | blank | Hand over only the medication this patient needs — never the whole med room. |
| Approval Mode | Manual week one, then Smart | See every action first; graduate to standing orders once they feel predictable. |
| Redact Secrets | ON | The privacy curtain. Non-negotiable. |
| File Checkpoints | ON | The "before photo" — lets you roll back if an edit goes wrong. |
| Sudo Password | blank — never store it | Don't leave the narcotics cabinet unlocked to save a few seconds. |
| Gateway | Local | Hermes runs on your own computer, close to you. |
Everything else on the pages that follow is context, so those rows make sense — and so nothing in Settings ever feels like a mystery switch again.
1Workspace & tool scope
This screen decides how Hermes may use your computer when it runs tools — terminal commands, file reads and writes, code. Think of it as the scope of Hermes' work area: when Hermes opens its toolbox, what room is it standing in?
"When Hermes opens its toolbox, what room is it standing in?"
This is the default folder Hermes uses for terminal commands and file work. If it's not set, Hermes uses whatever folder it happens to start in — which can be confusing, and lets it wander. Set it to one dedicated, safe, no-PHI folder, for example ~/My-Nurse-AI-OS.
Do not point it at your home folder, Desktop, Downloads, Documents, or / — those contain unrelated private files. One dedicated folder, and Hermes stays inside it.
"When Hermes runs code, should it stay inside the current project?"
Keep it on Project — the beginner-safe setting. It tells Hermes: work inside the project; don't treat the whole computer as the workspace. A stricter sandboxed mode, if your version offers one, is fine too. Avoid global or unrestricted modes.
"Does the terminal remember what the last command set up?"
Keep it ON. It means things like the current folder and environment carry over between commands — convenient, because setup steps don't have to repeat. The one catch: if something behaves strangely, the shell may be remembering old state. Start a new session, or ask Hermes to check where it is and what's active.
"Which of my computer's variables — including secret keys — get handed to tools?"
Leave it blank. Some environment variables are secrets (API keys, tokens). If you pass too many through, tools may access keys you never meant to expose. If a tool genuinely needs one later, pass only that specific variable by name — never * or "all".
"How much of a file can Hermes read in one request?"
The default (around 100000 characters — roughly 15,000–25,000 words) is right. It protects performance, your context window, and your privacy from accidental giant reads. Raise it only if Hermes repeatedly tells you a file was truncated and you intentionally want more.
2Models & reasoning
This screen chooses which AI model Hermes uses. At the top you pick the main brain; below that, whether smaller helper tasks use the same brain or a cheaper one. Note the fine print: changes here apply to new sessions — to switch the chat you're already in, use the model picker in the message composer.
"Who is the charge nurse for this conversation?"
The main model handles the actual work — planning, writing, coding, tool calls, decisions. This is the most important choice on the screen. Pick one strong, current model you trust and stay with it; the default Hermes offers is a fine starting point. Don't shop models daily — consistency beats novelty.
"How much thinking budget does the model get?"
Medium is right for normal serious work. Low is for quick answers and simple rewrites. Reach for High only when you'd say things like "take your time," "deep review," or "find what we're missing" — it's slower and costs more. For daily work, Medium is enough.
"Do the helper tasks get their own staff, or share the charge nurse?"
Behind the scenes, Hermes runs helper jobs — reading images, summarizing web pages, compressing long conversations, naming sessions. The default, auto · use main model, sends them all to your main model. Keep it that way while you're learning. It's the simplest, highest-quality setup. Don't over-optimize yet.
Later, if cost matters, downgrade only the disposable jobs — and never the safety-critical ones:
| Helper task | What it does | Cheaper model OK? |
|---|---|---|
| Title gen | Names your sessions | Yes — lowest risk |
| Skills hub | Searches available skills | Yes |
| Web extract | Summarizes web pages | Yes, if research volume grows |
| Vision | Reads your screenshots | Keep strong — especially for setup/security screenshots |
| Compression | Condenses long chats so context survives | No. Weak compression loses important decisions |
| Approval | Judges whether commands are safe to auto-run | No. This is part of the safety perimeter |
| MCP | Routes external tool calls | No — keep strong |
| Curator | Maintains your skills | Medium at least — a careless curator misfiles useful skills |
3Mixture of Agents
Mixture of Agents lets Hermes ask several AI models for their opinions, then have one stronger model act as the synthesizer that combines them into a final answer. Reference model 1 is one consultant, reference model 2 is another, and the aggregator is the lead nurse who writes the final organized note. The aggregator is the acting model — the one Hermes treats as in charge.
Big decisions worth a second and third opinion: strategy, governance review, "what are we missing?" checks, complex debugging, comparing options, high-stakes writing you're about to publish.
Routine work — quick questions, simple summaries, small rewrites, casual chat. You're paying several models instead of one, and waiting on all of them. It's boardroom mode, not daily mode.
Two practical notes. First, name your preset for its purpose — something like deep-review or stewardship-council instead of default, so future-you knows what it's for. Leave Context Window at 0 (auto-detect) unless you have a specific reason. Second, and more important:
Best rule: single model for routine work; Mixture of Agents for decisions worth a second and third opinion.
4Safety & approvals
For Nurse AI OS, this is the most important screen in Hermes. It controls how much Hermes may do automatically, when it must ask you first, how it handles secrets, and whether file edits can be rolled back.
"Which actions can proceed on standing orders, and which need a human signature?"
Start on Manual for your first week — you see and approve every action before it runs, which is how you learn what Hermes actually does (this matches the Setup Safety Checklist in your Starter Kit). Once its actions feel predictable, Smart becomes the right working default: low-risk commands proceed; risky ones — deleting files, install scripts, changing system settings, deploying, anything destructive — still ask you. Drop back to Manual any time you're handling secrets, doing server setup, working near private files, or simply feeling unsure. Never use fully-automatic / "YOLO" mode for anything involving your files, installs, credentials, or a public site.
"How long does Hermes wait for my yes or no?"
The default is often 60 seconds. If you multitask or step away — most of us do — set it to 120 so approval prompts don't expire while you're still reading them.
"Should Hermes ask before reloading its external tool connections?"
Keep ON. MCP servers can connect Hermes to files, databases, browsers, and online services — reloading can change what Hermes is able to reach. That should stay governed, with you in the loop.
"Which commands skip approval entirely?"
Leave it blank. Every command you allowlist is a command that bypasses your safety gate. In particular, never allowlist deletion, admin (sudo), download, install, or permission-changing commands — those have real side effects.
"Should detected passwords and keys be hidden from the model?"
Keep ON, always. This is non-negotiable. It hides API keys, tokens, and password-looking strings from model-visible content when possible.
"May Hermes reach addresses inside my own network, like localhost?"
Allow Private URLs, Browser Private URLs, and Local Browser For Private URLs control whether Hermes and its browser tool can reach private addresses — local dev servers, internal dashboards. If you're building and testing pages on your own machine, turn them on together (local browser is the safer pairing when private URLs are allowed). If you're not doing that kind of work, keep them OFF — less surface area is safer, and most beginners never need them.
"Does Hermes take a snapshot before it edits my files?"
Turn this ON — it often ships off, and it's the single biggest improvement you can make on this screen. Checkpoints are rollback snapshots taken before edits: if something goes wrong, you restore. For anyone whose folders hold real work, this is essential.
| Setting | Beginner-safe value |
|---|---|
| Approval Mode | Manual for week one → Smart (Manual when unsure) |
| Approval Timeout | 120 seconds |
| Confirm MCP Reloads | ON |
| Command Allowlist | blank |
| Redact Secrets | ON — always |
| Private URLs (all three) | OFF unless you build/test locally |
| File Checkpoints | ON |
5Voice
Two things live here: speech-to-text (you talk, Hermes types) and text-to-speech (Hermes reads answers aloud). Voice is wonderful for capturing a thought before you lose it — and it deserves the same discipline as anything else you say near a chart.
Turn Speech to Text on if you like dictating. If your Hermes offers a Local speech-to-text provider, choose it — transcription happens on your own machine instead of being sent to a cloud service. That's the stronger privacy choice. A Base-size local model is fine to start; step up to Small or Medium if it keeps missing words (larger models handle noise, accents, and medical vocabulary better). Set the transcription language explicitly so the model doesn't guess.
"Show me exactly what Hermes heard."
Keep ON, at least while you're new — Hermes posts the raw transcript into the chat so you can catch mishears before they become mistakes. Turn it off later if it clutters.
Keep OFF. Otherwise every answer is spoken out loud — distracting, and potentially private if others are nearby. Ask for audio when you want it instead.
A max recording length around 120 seconds is right. Better practice than long recordings: one voice note = one thought, one decision, or one task. Long rambles are harder for both of you to act on.
6Provider accounts
The "Connect an account" screen lists the AI providers Hermes may use, so you're not pasting keys every time. Some connect through a browser sign-in, some through an API key you paste once. The trash icon beside each one disconnects it — don't click it casually.
Fewer than you think. One solid provider is enough to start; two or three give you a fallback if one fails or gets expensive. A router-style provider that offers many models through one key is the most practical single addition once you're comfortable — useful for fallbacks, comparing models, and Mixture of Agents setups.
Resist overconnecting. Every provider you add is another account, another bill, another credential to protect, and another company that may see your prompts. Connect a provider when it has a real role, not because the button is there.
A connected provider is a possible destination for your prompts — not an automatic one. Content generally goes only to the model you select. But once connected, that provider's models can be chosen for main work, helper tasks, vision, and Mixture of Agents — so the perimeter still applies everywhere: no PHI, no patient screenshots, no employer-confidential files, no secrets pasted into prompts.
Use low-permission keys where possible, set spending limits with each provider, never store keys in public places, and rotate any key that may have been exposed. Watch for connectors whose fine print says a subscription needs extra usage credits — connecting it doesn't always mean it will work without them.
7Gateway connection
This screen answers one question: is Hermes running on this computer, or is this desktop app a remote control for a Hermes running somewhere else, like a rented server?
Hermes runs privately on your own machine, works offline, and your files stay local unless you explicitly send content to a model. This is the default, and the right choice for learning, personal files, and anything private.
The bedside workstation: private, close to the patient, works even when the network is down.
The desktop app becomes a front-end for a Hermes backend on another machine — useful for always-on jobs that run while your laptop sleeps. That backend is cloud infrastructure and must be treated like it: locked-down access, and no-PHI operational work only.
Central command: powerful for coordination, but it needs access control and stricter rules.
If your Hermes supports profiles, don't point all of them at a remote gateway. Keep personal and private profiles local, and give remote work its own dedicated profile — especially any profile tied to your employer, which should stay local and extra careful. Before saving a remote URL, use Test remote; prefer "Save for next restart" until you've confirmed the connection works. If a gateway won't connect, Open logs is where the answer lives.
8Tools & keys
The Tools & Keys pages are Hermes' supply room: optional service keys for web search, page crawling, media generation, voice, code publishing, memory, and observability. Each key you add gives Hermes a capability — and creates a credential to protect.
Only stock what you use. A useful working set for research and building is: one web search service, one page extraction/crawling service, a code platform connection if you publish, and a voice service if you use audio. Media generation is optional. Cloud-browser and automation services are advanced — leave them empty unless a real workflow demands them.
A memory provider gives Hermes continuity between sessions — preferences, project context, durable patterns. Valuable, with one discipline: memory is persistence, so never let PHI, secrets, patient stories, or employer-confidential material into it. One memory provider is enough; don't stack several.
Observability / tracing services record logs of model calls — prompts, outputs, costs. Powerful for debugging and evaluation later; unnecessary for beginners. If you ever enable one, know where traces are stored, set retention deliberately, and keep it no-PHI.
Ephemeral system prompts, prefill message files, messaging-channel settings, proxy URLs, custom endpoints: if you don't recognize a field, the safe setting is blank. These exist for specific advanced deployments, and the people running those deployments know they need them. Blank is not a gap — blank is a choice.
9Advanced limits & subagents
The Advanced screen sets limits on tool output, how many steps Hermes may take, and how delegated helper agents behave. The shipped defaults are well-balanced. Here's what each group means, so the numbers don't intimidate you.
Terminal output limit (~50000 characters), file page limit (~2000 lines), and line length limit (~2000 characters) all exist to stop huge logs and giant files from flooding a conversation. Keep them. Checkpoint limit (~20) is how many rollback snapshots are retained — plenty.
Max agent steps (~60) caps how many tool-using turns Hermes takes in one run — raise it only if long multi-step jobs keep stopping partway and you trust the workflow. API retries at 3 is right. Service tier: leave blank/default unless your provider account has a specific tier. Tool-use enforcement: keep auto — Hermes decides when tools are needed.
"Hermes' float pool: helpers it can send off to work in parallel."
Subagents are helper agents Hermes spawns for parallel work — one inspecting documents, one reviewing code, one researching, findings synthesized at the end. Sensible defaults: model and provider Not set (they inherit the main model — simplest), turn limit around 50, parallel subagents around 3, timeout around 600 seconds. Lower the parallel count if costs spike or your machine slows; pin a cheaper subagent model only once you know what you'd use it for.
If Hermes updates itself while there are local source edits, Stash preserves them and Discard throws them away. Keep Stash — it's the safer choice.
10Housekeeping
If unset, new sessions default to your entire home folder — a broader starting point than Hermes needs. Set it to your dedicated workspace, e.g. ~/My-Nurse-AI-OS, so every new session starts in a safe, organized place. (Yes — this is the same principle as the Working Directory in chapter 1. Assigned unit, always.)
Archiving hides a chat from the sidebar without deleting anything — every message is kept. Use it for finished experiments, completed setup threads, and stale conversations you might want again. On a Mac, ⌘-click a chat in the sidebar to archive it (Ctrl-click elsewhere).
The About screen shows your version and whether you're current. Automatic update checks are on by default — leave them. Release notes are worth a skim after updates, since settings screens move.
If you ever need to remove Hermes, there are three levels — always try them in this order:
| Option | What's removed | When |
|---|---|---|
| 1 · Uninstall Chat GUI only | Just the desktop app. Agent, config, chats, secrets all stay. | GUI misbehaving; reinstalling the app. Safest — start here. |
| 2 · GUI + agent, keep data | App and agent runtime. Config, chats, secrets kept for reinstall. | The agent itself is broken; you want a cleaner reinstall. |
| 3 · Uninstall everything | App, agent, and all data — config, chats, scheduled jobs, secrets, logs. | Only for a deliberate full wipe, with backups made first. |
Never click "Uninstall everything" casually. Your Hermes holds provider connections, keys, chats, memory, and project history. Level 3 takes all of it.
Reaching for uninstall because something's broken? Read When Things Go Wrong first — most problems resolve several rungs below a reinstall.
If you remember nothing else
Every screen above earns its place, but these five carry most of the safety. Check them once, and your Hermes is configured like a nurse set it up.
Configuration isn't a test you can fail — every setting here can be changed again tomorrow. Set the five above, leave the rest at their defaults, and get back to the real work: learning what Hermes can do beside you. When a screen confuses you, this handbook will be here, and so will the official docs.
Companions to this handbook: the one-page cheat sheet for install and first launch · the Hermes Essentials Guide (PDF) · the Hermes Power Guide (PDF) for profiles, scheduled jobs, and model routing · When Things Go Wrong, the troubleshooting companion for the day something breaks · and Start Here, which is still your actual setup path.
The safety perimeter
Do not enter patient names, charts, screenshots, identifiers, room numbers, employer secrets, passwords, account numbers, or patient-specific clinical details into Hermes, the quizzes, dashboards, or any AI tool — unless an approved institutional environment and accountable governance explicitly allow it. That perimeter holds in every chapter above: in prompts, in voice notes, in memory, in Mixture of Agents, on every provider.
Hermes can help with learning, organization, drafting, research support, project planning, and reflection. It does not carry a license, validate competency, approve clinical deployment, replace educators, or make patient-care decisions.
Agents propose. Humans judge. Nurses steward. Gray areas? Bring them to the monthly Lamp Huddle — a question asked early beats an incident reported late.