Your company's data is a foreign country.
Scattered across TripleWhale, Google Sheets, Obsidian, Discord and a dozen ad accounts. Metronous forges it into one living, queryable company brain — with a team of AI agents your team just talks to.
A wall of tools, or one box.
The whole stack runs self-hosted on a single VPS — 12 cores · 47 GB RAM · Docker Swarm behind HTTPS. A "building manager" keeps ~22 services alive and restarts anything that falls. $0 per query for the numbers.
Two halves that never lie to each other.
One half holds the exact numbers (Supabase / Postgres). The other holds the why (RAG-Anything — a knowledge graph over your whole vault). Ask "how much?" → it reads the table. Ask "why did it work?" → it recalls the note, and cites it.
Not one memory. Five.
A real brain remembers different things in different places. Metronous keeps five — and each one knows exactly what goes in it.
Numbers memory
Every ad, every day — the exact figures. Postgres.
Knowledge memory — the "why"
Research, winning copy DNA, angles, ICPs. A graph + embeddings, always cited.
Episodic memory
Every conversation & daily log — what was asked, said, decided. Timestamped.
Procedural memory — skills
The how-to's the agent learns, improves and reuses (more on this next).
Per-agent memory
Each AI employee keeps its own private notes, isolated from the others.
The brain refreshes itself.
A robot pulls fresh ad data from TripleWhale every 30 minutes and re-pulls a rolling 14-day window — so late conversions and refunds correct themselves. No human in the loop.
Three agents. Each its own mind.
One runtime (Hermes) hosts three independent AI employees — different models, different jobs, each with its own Discord bot, skills and memory, all reading one brain.
A soul, and a set of skills.
The model is just the engine. What turns it into a real employee are two things: a SOUL and Skills.
Who it is + the rules
A written manual injected at the top of every single conversation. Plain text, not code.
- Personality & brand voice
- The thinking frameworks (e.g. DRIVE, 20 angles, 6 ICPs)
- Hard rules — FTC/FDA compliance, no jailbreak
- The anti-hallucination contract
Playbooks it pulls only when needed
Reusable how-to's. The agent opens just the one the task calls for — so it stays focused.
- Agents can write new skills from experience
- …and self-improve the ones they have
Every use leaves it smarter.
This isn't a static tool — it's a recursive intelligence. The more the team uses it, the more it learns: outcomes become memory, repeated work becomes new skills, and winners feed the next round.
with use
Winners get detected, ported and reused. Skills get graded and consolidated. Memory only grows. The flywheel turns on its own.
The whole business, one glance.
The morning board. One screen answers "how are we doing?": every KPI with a period-over-period delta, the daily sales-vs-spend curve, who's winning, and what the AI team has been up to — all reconciled to the exact same source as Tracking.
A P&L for every strategist.
Tracking auto-decodes the cryptic BATCH# ad-name code into the real person — then gives each one their own P&L. Click a row → that strategist's exact ads and creatives, with video playback. Sortable, date-filterable, no manual tagging.
| STRATEGIST | ADS | RESULT |
|---|---|---|
| Chester | 2,129 | |
| Alya | 1,749 | |
| Arcie | 30 | |
| Gerry | 136 | |
| Dell | 134 |
▸ click row → ads · creatives · video
From winning DNA to a new ad.
The ad factory. A seeded company brain — avatars, angles, winning copy — feeds a generation engine that writes fresh copy with Claude Opus 4.8 and edits the image with fal.ai. Turn a proven winner into new variants, on tap.
fal.ai image
The cockpit for your AI copywriters.
Pick a copywriter persona, choose its model, and drive it — all from one panel. It shares the Static Production brain, so every line is grounded in real winners. Personas can dial intensity up ("Little d") while the FTC/FDA compliance floor never moves.
Ask in plain words. Get cited answers.
Numbers come back from SQL; the "why" comes back from the brain — with citations. If it doesn't know, it says so. Chat is even multimodal: send an image or a video frame, get a generated image back.
Not a demo dataset. A real one.
Every screen is built on real, live-ingested data — refreshed automatically.
Every number comes with a receipt.
Numbers always come from SQL — never invented. The reasoning comes with citations. And when the data isn't there, it says "I don't know." Anti-hallucination by design.
Numbers ← SQL
Every metric is queried, never guessed.
Why ← cited recall
The reasoning shows its sources + dates.
Unknown → "I don't know"
No filler. No hallucinated answers.
HOSTED
$0/q