Overview / Hermes & Asset Ops

Hermes · always-on assistant

An assistant that does the busywork — with a hand on the brake.

Hermes runs around the clock on VPS2. It's organised into profiles — each a persona with its own channels, scheduled jobs and skills. Nothing important happens without a human tap: actions are proposed, you approve them from Telegram. Two profiles do the heavy lifting: huxberrypm and Asset Ops.

profiles gateways cron workflows skills & plugins human-in-the-loop

The runtime

Channels in. Actions out.

Hermes listens on your channels, runs scheduled jobs, and proposes actions you approve — turning messages and meetings into tracked work. Hover any node to trace it.

Channels Hermes What it produces

Channels

Telegram Gmail Google Meet Fizzy

↓ Hermes proposes

Hermes

↓ you approve

What it produces

Fizzy cards Daily brief Directus DAM GBrain

Profile

A scoped persona — its identity, channels, jobs and skills. huxberrypm is one; assetops is another.

Gateway

How Hermes is reachable: Telegram, Gmail, and more. Messages in, approvals back.

Cron workflow

A deterministic job on a schedule — scan, summarise, propose. No free-wheeling.

Approval

Proposals wait for a tap. cron_mode: deny means scheduled jobs never auto-act.

01 — The profiles

What each profile actually does.

profile · huxberrypm

The PM assistant

A proactive chief-of-staff for meetings, email and tasks — it watches the day and proposes the next move.

Telegram Gmail

Reads & connects

Google MeetCalendar Fizzy boardsClickUp · readOdoo · read

Runs on a schedule

Meeting notes → proposed Fizzy task cards · hourly
Gmail action-markers → proposed cards · hourly
Correspondence → contacts wiki · hourly
Daily brief email, morning & afternoon · 7:30 & 15:00
Critical-email follow-up sweeps · late afternoon

Safety

cron_mode: deny — every scheduled job proposes; you approve or reject from Telegram inline buttons (handled by the nish-fizzy-telegram plugin).

profile · assetops

The asset pipeline

Turns a folder of raw supplier photos and videos into a described, tagged, searchable asset library — in hours, not weeks.

Gemini vision Directus DAM

Writes to

Directus DAMR2 · originalsCloudflare Stream · video

How it's triggered

Run per supplier from the command line on VPS2
Cost guard estimates spend; a second gate fires over a threshold
Idempotent — safe to re-run; existing assets are updated, not duplicated

Status

The engine runs today as a CLI pipeline on VPS2. An optional Telegram approval layer exists but was retired in May 2026 in favour of direct runs.

02 — The Asset Ops pipeline

From a Drive folder to a shoppable library.

Five phases. Only one of them writes anything — the rest are safe to run and re-run while you watch the cost.

01 · discover

Find the real subjects

Scan the supplier's Drive folder and dedupe to canonical subjects — no double-processing the same product.

read-only

→

02 · prepare

Download & resize

Pull the high-res originals, make working copies, cache locally for analysis.

read-only

→

03 · analyse

Gemini describes it

Photos get a 14-field description — subject, tags, alt-text, suggested uses. Videos get a timestamped shot list, reusable segments and key quotes.

read-only

→

04 · publish

Into the DAM

Originals to R2, video to Cloudflare Stream, and typed rows (with all the metadata) into Directus.

writes

→

05 · verify

Reconcile

Every DAM record is checked against its stored file and analysis artifact. Nothing dangles.

read-only

The payoff. A marketing or design team member opens the DAM at dam.huxapps.com and finds every asset already described and ready to use — searchable from day one, instead of a shared drive nobody can navigate. Each asset arrives:

tagged captioned alt-texted labelled for where it can be used

03 — Where the library goes next

A described library is a generative one.

The point of describing every asset isn't only search. Once the library knows what it holds — down to the second — it becomes raw material the company can generate new marketing from, instead of shooting everything from scratch.

The library knows itself

Every photo carries a 14-field description; every video is broken into a timestamped shot list. Ask for "the slow pan across the walnut headboard" and the system knows the clip — and the exact seconds it lives at.

Generate, don't stock-shoot

The goal: feed those described assets into Higgs Field AI for video and image generation — so new marketing is built from our own products and rooms, staying on-brand instead of generic.

Long-form, on tap

Reusable segments and key quotes are already catalogued. Stitch them into long-form video and fresh imagery for a campaign — far less shooting, far faster turnaround.

Where we're headed. The Higgs Field AI integration is the next step on the roadmap — not live yet. But the second-level tagging that makes it possible is being built into the pipeline today.

04 — What Hermes touches

One assistant, many hands.

TelegramGmailGoogle Meet Google CalendarFizzyClickUp OdooDirectus DAMGBrain Cloudflare R2Cloudflare StreamAgent Vault