Mark-1 and the stock Hermes Agent share the same codebase. Over the past weeks they have evolved in different directions — Hermes hardens a general-purpose platform for everyone; Mark-1 has been built upward into a single executive's operating system. This is the map of that divergence, measured exactly.
Source of truth: the public NousResearch/hermes-agent repo on this VPS · pinned at v2026.6.5 + 301 commits · HEAD 4219a91df · all figures below counted directly from disk.
Both systems run the identical Hermes Agent engine (v2026.6.5, the v0.16.0 line). What separates them is not the engine — it is what each is being optimised for. Hermes optimises for breadth across millions of unknown users: it ships 75 core skills enabled by default plus 95 optional ones you can switch on. Mark-1 optimises for depth for one known operator and one business: it keeps every relevant upstream skill and adds 104 of its own.
Read plainly: upstream Hermes is getting more reliable and broader. Mark-1 is getting more yours. Both matter — Mark-1 inherits every upstream stability gain on each safe update, then layers its own vertical stack on top.
Not all 193 are bespoke — and claiming so would be dishonest. Here is the exact composition, counted by set intersection against the upstream repo. 89 skills are inherited verbatim from Hermes; 104 were written specifically for Mark-1.
For symmetry: upstream also ships 81 optional skills Mark-1 has not enabled (blockchain, large parts of the mlops catalogue, finance widgets, etc.) — capability that exists in the platform but isn't relevant to this operator. Breadth is not the same as fit.
From a shared origin, each system's recent work pulls in a different direction. Left: what upstream actually shipped (verifiable in the git log). Right: what was built into Mark-1 over the same window.
The asymmetry is the message. Upstream's column is fixes to the shared engine. Mark-1's column is new organs grown on top of it. Neither replaces the other — Mark-1 keeps pulling the engine fixes in via safe updates (the hermes-safe-update skill exists precisely for this).
These are skill domains absent from the upstream repo entirely — built specifically for Mark-1. For each headline subsystem: what it is, why it was implemented, how it benefits you, and exactly what stock Hermes does in its place.
To turn the agent into a sellable "AI for Executives" SaaS at $980/mo — a real revenue line for Nionium, not an internal tool.
A commercial asset you own end-to-end: chat UI, Stripe billing, invite codes, voice mode, admin dashboard, per-tenant isolation. Mark-1 is both the operator and the product.
Upstream's memory is a flat string injected per turn. An executive needs a connected, queryable knowledge base — pages, entities, relationships — that compounds across sessions.
Ask "what do we know about X" and get a synthesised, cited answer from everything ever captured — with an honest "gaps" section. Cross-references your goals automatically; self-maintains nightly.
You shouldn't have to ask. The agent should ingest feeds daily, reason about them against your strategic goals, and surface only what creates or removes an opportunity or risk.
An analyst that works while you sleep — AM sweep + PM watch, tiered alerts (urgent / digest / silent), strictly inside a hard money boundary it can never cross without you.
Geopolitics (Iran, Ukraine, Georgia) directly affects your portfolio and decisions. You wanted a live theater map and a daily intelligence product, not a news feed.
A self-updating military-standard map (forces, flights, vessels, pipelines, cables) plus an original daily SITREP in Mark-1's own analytical voice — published, archived, shareable.
Unattended runs were either grinding to the turn cap or reporting "done" when nothing actually changed. The loop needed to stop for the right reason and prove its own work.
Autonomous jobs self-correct: they halt when stuck (and say why), and refuse to claim success until reality is independently verified. Fewer false "done"s reaching you, zero new approval prompts.
Mark-1 must maintain itself: pull upstream fixes without losing the vertical stack, tune its own gateway, clone instances, debug its own TUI, register custom providers.
The meta-layer that keeps Mark-1 maintainable. This is precisely how Mark-1 inherits all 301 upstream commits while keeping its 104 bespoke skills intact — no fork drift, no lost work.
You want true sub-second voice interaction (and a live avatar) tuned for an executive persona — not the generic voice path the platform ships.
Hold real-time voice conversations in the Marin voice with three-layer noise suppression, plus an optional live video avatar — desk or pocket.
You operate from your phone via Telegram. Running multiple Mark-1 instances on one bot token required solving polling-conflict ownership and webhook routing.
Mark-1 reaches you reliably on Telegram (with Mari read-only), delivers the morning briefing there, and the hub-vs-spoke ownership model is documented so spokes never hijack the bot.
Beyond these subsystems, the remaining net-new skills are 45 data-source integrations and content workflows — detailed in the next section — the breadth that lets Mark-1 answer from real data instead of generic knowledge.
Stock Hermes ships generic search and a browser. Mark-1 is wired directly into authoritative live data sources, so it answers from the actual numbers — markets, filings, science, geodata — not from model memory. This is the data hierarchy in practice: check the integrated source first, web search only as fallback.
Each row maps to a verified integration skill in ~/.hermes/skills with a working key or auth path. This is the practical difference between "an agent that can search the web" and "an agent plugged into the world's authoritative data feeds" — the latter is what lets Mark-1 show you the real series, the real filing, the real price.
The same dimensions, both systems. Every cell is a fact, not a slogan. A cross (×) on the Hermes side is not a criticism — a general platform should not ship your business context.
| Dimension | Stock Hermes Agent | Mark-1 |
|---|---|---|
| Skills available | 75 core + 95 optional = 170 | 193 active (89 inherited + 104 net-new) |
| Persistent context | Flat memory string per turn | Memory + Second Brain graph + user profile, applied unprompted |
| Autonomy | Cron tool exists; reactive use | 20 scheduled jobs: sweeps, briefings, SITREPs, backups, watchdogs |
| Self-correction | max_turns ceiling only | Goal-aware stop + mandatory verification |
| Product surface | None — it is infrastructure | DataDesk SaaS: chat, voice, billing, admin |
| Knowledge synthesis | Answers from context window | Hybrid retrieval + cited synthesis with honest gaps |
| Intelligence layer | None | Feeds reasoned against goals; tiered signals; Intel Map |
| Business awareness | Generic, anonymous | Knows portfolio, properties, family, Nionium goals |
| Deployment surface | Local CLI / gateway | 25 live project artifacts + multi-VPS tenancy |
| External data integrations | Generic web search + browser | 36 live APIs — finance, science, geodata, comms (source-first) |
| Codebase | Identical — both run Hermes v2026.6.5 (HEAD 4219a91df). Mark-1 is upstream + a curated skill/subsystem layer, re-synced via the safe-update skill. | |
The concrete work behind the divergence — each a capability that did not exist in Mark-1 before, with the reason it was built.
Unattended jobs now halt for the right reason and prove their own output against reality before reporting. Closes the "reported done but nothing changed" failure class. Zero new approval prompts for you.
A self-verifying cron that reclaimed 253 GB and keeps the production VPS from filling with build cache — with a built-in container-count safety check so nothing is ever damaged.
7am Tbilisi executive briefing — unread mail, calendar, tasks, docs-to-sign, international + Iran-theater news — delivered as DataDesk cards and Telegram, with sound + desktop alert.
Aggregated Telegram / RSS / YouTube panel. Solved the EU-region YouTube block via the innerTube WEB-client path — a genuinely hard problem the public API refuses.
Real money can flow: live checkout, signed webhooks, $0 invite-code activation. DataDesk became a real commercial product end-to-end.
Per-tenant isolated VPS model, relay hub, HMAC-signed Brain calls, user-scoped data — the path to selling DataDesk to more than one executive.
Hybrid retrieval with inline citations and an honest gaps section, wired into the chat with a sources panel — durable knowledge that compounds.
This report does not estimate. All counts were taken directly from the upstream repo checked out on this VPS and Mark-1's live skill directory, by exact set operations.
git remote -v → github.com/NousResearch/hermes-agent; git describe --tags → v2026.6.5-301-g4219a91df.find /usr/local/lib/hermes-agent/skills -name SKILL.md → 75 core; same over optional-skills/ → 95 optional.find ~/.hermes/skills -name SKILL.md → 193.~/.hermes/skills but absent from both upstream skill trees → 16.git rev-list --count v2026.6.5..HEAD → 301 commits, inspected by message for the hardening examples shown.Correction note: an earlier draft of this page reported "+157%" by comparing Mark-1's 193 against upstream's 75 core skills only — ignoring the 95 optional skills upstream also ships. That figure was withdrawn. The honest comparison is 170 available upstream vs 193 active in Mark-1, of which 104 are genuinely net-new.
The upstream agent will keep getting safer and wider — and Mark-1 inherits every one of those gains on each safe update. But the divergence is permanent and intentional: Mark-1 is no longer "a Hermes install." It is an executive operating system that happens to run on Hermes, with 16 capability domains and 104 net-new skills the platform was never meant to carry.