Open-source · running a real company

A video call became a P&L. Same day.

A refund response drafted in brand voice in 90 seconds. A six-bank liquidity view refreshed every six hours. 116 signals compressed into ten founder decisions in one triage run. This is the AI operating system we run our 8-figure company on — €21 a day, built without an internal engineering department.

8-figure revenue +100% YoY 4,842 Brain docs 7 agents + command center 373 available skills 97 MCP tools
The unlock

AI that knows how your company actually works.

Your company already knows almost everything it needs. It just can’t remember it — because it was never captured, never digested. Any model can reason; none of them know your refund policy, your supplier terms, or what you decided last Tuesday. We turned approved operating context into a living brain: what the company knows (4,842 documents), what it knows how to do (373 available skills, 47 canonical), and authenticated hands in its systems (97 MCP tools). The AI reads it before it acts, checks the source, and writes back a receipt after consequential work. Context in. Verified execution out.

Architecture

Capture is a storage unit. A brain is five layers.

Everyone dumps their docs into a vector database and calls it a company brain. That's a storage unit — necessary, but raw material doesn't decide anything. The value is the intelligence layer on top. We learned it the hard way: we started at capture, poured everything in, and three months later memory itself was the bottleneck — thousands of documents, retrieval degrading, the system technically smarter but operationally messier. So we rebuilt around five layers.

01 · CAPTURE
The raw material
Slack, email, meetings, documents and the systems that run the business — recorded as they happen, stamped with source, time and sensitivity. When our extraction model went down for two days, we lost nothing: capture never depends on the model being up. The capture spine →
02 · RETRIEVAL
The operating layer
Not the whole company history — the handful of things that matter for the task in front of it, promoted onto domain shelves. The finance agent searching “returns” hits the refund policy, not fabric returns to a supplier. Domain indexes →
03 · SOURCE OF TRUTH
Which source wins
When the live system, an old SOP and a Slack correction disagree, the conflict is flagged, a human adjudicates, and the loser is marked superseded — never silently overwritten. Agents stop being confident liars with better formatting.
04 · PERMISSIONS
The right brain for the workflow
Every caller is authenticated today. Brain Spaces is the deny-by-default retrieval contract, with fine-grained sensitive-domain scoping rolling out domain by domain. We publish the boundary and the remaining rollout instead of treating a design as universal deployment. Brain Spaces →
05 · FEEDBACK
The layer that compounds
Every human correction becomes a rule. And the system watches its own queues, not totals — a stalled organ opens a task instead of hiding behind a green dashboard. Organ health →
→ EXECUTION
Agents and people, acting on it
Signals become proposed work with an owner, an authority level and a verification check. Four things always stop at a human: customer-facing, financial, people, irreversible. Every consequential action leaves a receipt. The context-to-work loop →
signalcandidateproposed workhuman yes / noreceipt

It's model-agnostic by design — plain files plus MCP. Swap the model underneath (Claude today, Codex tomorrow, whatever's best next year) and the company veteran stays. The brain is the asset. The model is just the worker that reads it.

The companies that win won't have the biggest prompt library. They'll have the cleanest intelligence layer. The model is rented. These five layers are yours.

In production

What formalized context unlocks.

Not demos. Things that happened inside the company, on real data, with real money attached.

video call predictive P&L

A video call became a P&L.

A retail partner pitched a pop-up over video. From the call material and a stock spreadsheet, the brain pulled our sales history and modeled footfall, conversion, commission and sell-through — a predictive P&L that answered their terms the same day.

zero code live treasury

The CFO ships her own tools.

Our finance lead — no engineering background — built a liquidity dashboard that reconciles six bank accounts against upcoming invoices. It refreshes every six hours. Nobody wrote a spec.

approved ticket history grounded CS

Support learned from the cases we approved.

The CS agent uses an approved ticket corpus, a governed knowledge base and decision rules to answer routine cases and prepare higher-risk responses for review. New conversations improve the system through the same source and privacy controls.

423 meeting notes searchable memory

Meetings became memory.

Generated notes from approved meetings are captured, deduplicated and source-linked. Native transcript coverage is a known gap, so we claim searchable notes rather than pretending every spoken word is indexed.

The human layer

One team. One operating memory.

Agents cover the recurring work. Everything else runs through the team — every department on the same AI client, over the same memory. Same knowledge, same rules, same answers. The brain isn't an agent feature; it's how the company executes.

the founder · monday 9:00

"What changed last week?"

Six minutes later the weekly brief drops in Slack — drawn from the agents' Sunday runs, the meeting notes, and the world model. No tabs opened.

the finance lead · friday 16:00

"Reconcile the week. What's off?"

Accounting, six banks and Shopify checked in one pass. Two anomalies flagged with the receipts attached. The weekend starts on time.

the cs lead · tuesday 11:00

"Why is order 8341's refund late?"

Full thread, policy match, and a brand-voice draft in 90 seconds. The decision writes back to the brain for the next similar ticket.

the merch lead · tuesday 10:00

"Sell-through by size. Anything to act on?"

Three SKUs flagged for transfer, one for markdown. Decision made before lunch — and Marketing's agent sees it in its next run.

None of these moments started with a dashboard or a tab switch. They started with a sentence — to an AI that already knows the company. That's the execution layer.

The math

Every seat. Every hour. 16:1.

The cost side counts all five Claude Max seats — the line most ROI math hides. The value side counts both layers: what the agents absorb, and what the team reclaims working over the brain. Conservative on both.

16.2:1
every €1 in returns €16 in labor
107h/wk
reclaimed · ≈2.7 full-time people
€21/day
runs the entire AI layer
~23 days
payback on the full stack
cost — 5× Claude Max (founder + every dept): €460/mo · Codex, API, Hetzner, Mac Mini, Tailscale: €171/mo€631/mo all-in

value — the agents absorb: 62h/wk × 52w = €77,584 / yr
value — the team reclaims: 45h/wk × 48w × ~30 users = €45,360 / yr (≈90 min/person/week — meeting memory alone clears it)

total: €122,944 / yr on €7,572 → ratio 16.2:1 · payback ~23 days · cost ≈ 22% of one ops hire, output ≈ 2.7
No revenue impact claimed — every number is a cost avoided, every assumption in chapter 12. Audit the math →
Open by default

Fork it. Build your own.

This isn't a SaaS and it isn't a pitch. It's the actual blueprint — written inside a company that runs on it. Meant to be copied, not admired. And all of it is free: the playbook, the skills, the kit. No email gate, nothing to buy.

What this is

What it is, and what it isn't.

€21 a day runs 107 hours a week of company work — measured inside a real 8-figure company, with zero engineers on staff. Everything on this site is the how.

Not a SaaS.Nothing to buy, nothing to log into. A repo, a playbook, and a live dashboard you can audit.
Not a chatbot.Those tell you to build something. This tells you what to build — and ships the patterns to do it.
Not "AI with faces."No personality layer over an LLM. We document an operating system.
Not a kit.The hard part isn't the code — it's the discipline behind it. The playbook teaches the discipline.
Talk

Questions? Write me.

Stuck on the playbook, want feedback on what you're building, or want help wiring this into your company — I read every email. hello@usecompai.com