Personal Agent · What we add
Every DeepSenz Personal Agent runs an unmodified, pinned release of the open-source Hermes Agent. Our enhancements sit beside it—in configuration, profiles, skills and companion services—never inside its code.
The reference setup runs on Herman, the founder's Mac Studio in Da Nang. Product badges below distinguish what is available, in beta, and still in development.
How it fits together
Hermes stays the unmodified agent core. BrainBridge routes model work, MemSenz keeps approved memory, and CyberSENZ Lite performs scheduled checks of assets you authorize. Deep Recall is the separate life-evidence vault in development.
LM Studio can run the local model on your Mac. Optional private access has two distinct routes: LM Link reaches models across your devices; your own Tailscale network can reach the agent from your iPhone.
Your Mac · owner controlled
DeepSenz modules · outside Hermes code
Side by side
Same Hermes underneath. Different defaults, and a set of tested additions for privacy, voice and memory.
| Area | Stock Hermes | DeepSenz Personal Agent |
|---|---|---|
| Agent code | Hermes Agent, MIT License | The same code, pinned to a release we've tested on Macs |
| Main model | Whatever provider you set up | A local model on your Mac (Gemma 4 26B A4B by default); cloud only when you approve it |
| Background tasks | Helper tasks can discover and use available providers automatically | Every helper task pinned to your local model, with no automatic cloud fallback |
| Voice | Chained voice; the default voice output is a cloud service; GPT-Live on desktop | Fully on-device voice by default, plus full-duplex GPT-Live and a Gemini Live option |
| Memory | Two small notes files (about 1,300 tokens) plus searchable chat history | All of that, plus optional MemSenz owner-approved Markdown memory (Mac beta); Deep Recall evidence vault is in development |
| Model routing | One model per profile; you choose | BrainBridge and a local-first route policy: personal data stays on local models |
| Specialists | Bot Mode profiles you create yourself | A fast local helper plus an optional council of cloud specialists that stays idle until you choose one |
| API keys | Stored in a local settings file | Fetched at runtime from a secrets manager; never copied into profiles |
| Email and calendar | Google Workspace skill available | Read-only Gmail and Calendar, limited to two read scopes |
| iPhone access | Set it up yourself | Optional private agent access on your own Tailscale network; LM Link separately reaches models across devices |
The enhancements
Each card shows what the enhancement does, which Hermes front door it uses, and where it stands today.
Running on Herman
Your agent thinks with Gemma 4 26B A4B running in LM Studio on your Mac. A second "Fast" profile uses the smaller Gemma 4 E4B for quick answers. Neither has cloud keys.
Every background helper task—summaries, titles, memory upkeep—is pinned to the same local model with an empty fallback list, so nothing quietly calls a paid service.
Uses: Hermes config, profiles and its SOUL instructions, which require your approval before any paid cloud handoff.
Available now
BrainBridge is DeepSenz's thin routing layer with one job: send each task to the right model. A route policy decides first. Personal and sensitive data can only go to local models (LM Studio or Ollama). Cloud routes are off by default and, when you enable them, carry budgets, consent records and receipts.
Uses: a separate companion service that calls Hermes and model servers through their standard interfaces. It adds no code to Hermes.
Running on Herman
Beside the local agents, optional specialist profiles can reach frontier cloud models—Claude, Grok, Nemotron and GPT—through your own provider accounts. They stay idle until you pick one. Their keys are fetched at runtime from a secrets manager and never written into profile files.
Uses: Hermes Bot Mode and profiles.
Running on Herman
Speech recognition runs on local Whisper and replies are spoken by Piper, a local voice engine—so a spoken conversation never leaves your Mac and costs nothing per minute. You can interrupt mid-reply.
Measured on Herman: Piper rendered 3.3 seconds of speech in 1.2 seconds, and Whisper transcribed a phrase in under one second once loaded.
Uses: Hermes' chained voice mode with local speech providers.
Hermes built-in
For natural, talk-over-each-other conversation, switch to GPT-Live. OpenAI's live voice model handles the talking while Hermes does the real work with your local model and tools. OpenAI bills it to your own account, listed at $0.05 per minute of session time.
Uses: Hermes' desktop voice-engine setting. Microphone audio goes to OpenAI during a live session.
Pilot
A DeepSenz voice loop that listens through Google's Gemini Live and answers in a local voice. It wakes on a name, accepts follow-ups for 45 seconds, and closes the cloud connection after 3 minutes of silence so usage doesn't run on. If Gemini Live drops, it falls back to Gemini's standard API.
Uses: a separate companion service with local speech output. Gemini usage is billed to your Google account.
Hermes built-in
Hermes keeps two compact notes—one about your work, one about you—plus full-text search across past conversations. DeepSenz keeps all of it and can switch on review-before-save, so you approve what your agent remembers.
Uses: Hermes' own memory files and settings.
Mac beta · free test
An owner-approved, Karpathy-inspired Markdown wiki for facts, decisions, projects and preferences. New suggestions wait in a review inbox. You can read and edit approved pages and inspect local memory receipts.
Uses: a standalone local module plus a Hermes skill; no resident memory daemon. Join the free Mac beta →
In development Optional
For people who want their agent to remember a life, not just a chat. Today, the v0.1 local toolkit can ingest and verify evidence. The customer vault, owner-approved handoff into MemSenz and controlled recall are planned releases. Your agent will receive only memories you approve; raw evidence stays closed by default.
Evidence is preserved. Memory is derived. Knowledge is distilled. The current toolkit uses no model; future vault processing is designed for a local model only. Explore Deep Recall and its release plan →
Uses: separate local tooling; works with storage you choose, such as an external drive or home NAS.
Running on Herman
Your agent can read Gmail and Google Calendar to brief you, using only the two read scopes. It can't send, delete or change anything, and private content isn't passed to cloud specialists without your approval.
Uses: a minimal sign-in helper of our own, kept outside Hermes' code.
Pilot
DeepSenz Genies such as Travel Genie arrive as checksum-verified capsules and plug into your agent as a standard Hermes skill, so a specialist can join without changing the agent itself.
Uses: Hermes' skills folder.
Running on Herman
Before any Hermes update, your agent's state is backed up and verified. If an update misbehaves, we restore the backup together with the matching code and app build.
Uses: Hermes' official updater and its backup.
Beyond built-in modules
SkillSenz.com lists four downloadable free previews and 296 unreviewed publisher-source skills across 12 categories for inspection. None has earned a full SkillSenz Verified safety badge yet. Experience Capsules (CAPs) are the planned richer tier: standard skills with persona, dated knowledge, and experience patterns.
For agent developers, SkillSenz.ai mirrors the skills catalog in agent-readable JSON. OIP remains a separate technical reference. Your agent can inspect the published facts; installation, permissions, and any future purchase stay under your control.
Personal Agent
Choose the extra depth your Personal Agent needs. Sovereign is a complete private system, commissioned remotely with live human care.
Available now
Local-first model routing with optional cloud lanes under your control. $79 introductory one-time price for DIY; included with Core and Signature installations, separate from the $199 DIY Kit.
Available now
Scheduled read-only checks for approved websites and TLS certificates, with urgent alerts and a daily local report. $9.99 per month or $99 for 12 months after activation; the first 60 days are included with Signature installation. Managed CyberSENZ is planned.
Mac beta · free test
Approved, editable Markdown memory with a review inbox and local status light. The $49 one-time purchase is not open; early testers receive the test download free in exchange for feedback. Explore MemSenz →
In development
Optional, local life-memory vault in development. See the full vision and build status →
Premium
Frontier AI on two Mac Studios in your own home or office, commissioned with you remotely and supported live by a person.
From $59,000 (estimated)
Compatibility with Hermes upgrades
Because DeepSenz never changes Hermes' code, a new Hermes release installs with the official updater. We test our enhancements against it first, then recommend it.
Proof on Herman · September 2026
Herman jumped 2,451 commits to a new Hermes release in one step using the official updater. Its settings migrated from version 31 to 46 automatically, a verified backup was taken first, and DeepSenz add-ons—task router, voice bridge, memory retrieval and research bridge—passed their checks afterward.
We follow the standard release cycle rather than a private fork.
Voice, memory, routing, skills, permissions and rollback are checked on the reference Mac.
Exact versions and checksums, with notes on anything that changed.
With a verified backup first and a clear way back.
| Enhancement type | On upgrade |
|---|---|
| Settings and profiles | Migrated by Hermes' official updater |
| Skills and Genies | Re-tested; reinstalled through Hermes' standard skill paths |
| Companion services | Unaffected; they use stable interfaces, not Hermes internals |
| Your memory and archive | Untouched, and backed up before the update |
Optional setup · no DeepSenz charge
PPC gives your devices private ways to reach your Mac without opening public ports.
Your own Tailscale network can connect your iPhone to the agent. LM Link separately connects LM Studio models across your devices; it uses a Tailscale-backed network managed by LM Link and cannot join your existing tailnet. Both routes are optional and need to be configured and tested for your devices. How LM Link differs from a Tailscale tailnet →
Standard Hermes. Tested additions.
Credits and sources · checked October 2026
