Open-source AI operating system used by a real 8-figure company growing +100%.
Every recurring decision in our business used to live in someone's head. Refund policies in old Slack threads. Pricing exceptions as tribal knowledge. Incident playbooks as the founder's memory.
We extracted all of it. Structured it. Made it executable.
Pulled from email, Slack, support tickets, reports, meeting notes, and 18 months of decisions. Structured. Versioned. Searchable.
Procedures the brain can run on demand. Close the books. Triage a refund. Run weekly P&L. Documented, parameterized, callable.
Every business system the brain can read or write — e-commerce, helpdesk, accounting, marketing, analytics, warehouse. Governed via ACL. Audited via logs.
Every decision, fix, and pattern adds to the brain. Month six is not month one. The gap keeps widening.
Y Combinator's Tom Blomfield calls this "the missing layer between company data and reliable AI automation. Every company in the world is going to need one." We call it the brain.
Most "company AI" projects build a knowledge base. We built a memory that captures work where it happens, learns from it, and writes back what it learned. Bidirectional. Self-refreshing. With privacy hard-stops.
Slack threads, Meet/Gemini notes, Gmail, Google Drive. 228 meeting docs auto-captured, 1,019 signals promoted. Privacy hard-stops on HR, payroll, family, health.
Agents and employees write back to the brain when they learn, finish, or break something. Every interaction grows the memory. Notion AI reads; this writes.
Raw input → signal → task → output → decision → world model. Every layer has its own lifecycle. Health checks tell you what's broken before you notice.
Current-state, customer-signal, capabilities, gaps, DRI map. Refreshes weekly from real signals. Your AI knows what's true today, not 6 months ago.
Master prompt, MCP tools and brain access deployed in ~2 minutes per employee.
curl -fsSL https://your-mcp.example/setup.sh | bashirm https://your-mcp.example/setup.ps1 | iexTemplate in the playbook. Adapt to your MCP endpoint, your domains, your master prompt. Full implementation typically takes 6–8 weeks for a consumer SME — talk to us if you want help.
The brain is the substrate. What turns it into an operating system is what runs on top: eight AI agents 24/7, forty integrated tools, and every team member with one command and one shared master prompt. This is the pattern any consumer SME can replicate.
Eight domain agents run 24/7 on the same brain. Read context, execute, write back what they learned. All share one memory — the CS agent knows what Marketing decided last week.
44 MCP tools wire e-commerce, helpdesk, accounting, marketing, analytics, warehouse, and Slack to the brain. Every read and write logged in the action ledger. Auth observed.
Every team member connects in one command. Same master prompt as the agents. The brain captures their work back from Slack, Meet, Gmail, Drive — no documentation hour, no manual sync.
The LLM is the execution layer of the company. Tools write data. Agents write learnings. Employees write decisions. The brain compounds with every interaction — and that compounding is why month six is qualitatively different from month one.
On a typical Tuesday, dozens of operational decisions get made across customer service, retail, finance, marketing, and merchandising. None of them start with a spreadsheet. They start with a question — to an LLM that already has the context.
"What changed last week and what should I look at first?" 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.
A customer asks why their refund is late. "Order 8341 — timeline, who promised what, what should I say?" Full thread, policy match, and brand-voice draft in 90 seconds. The decision is captured for the next similar ticket.
"Reconcile last week. What's off?" Accounting, bank, and Shopify checked in one pass. Two anomalies flagged with the receipts. Weekend starts on time.
"Sell-through by size. Anything to act on?" Three SKUs flagged for transfer, one for markdown. Decision made before lunch. The Marketing agent sees the markdown plan in its next run.
None of these moments started with a spreadsheet, a query, or a tab switch. They started with a sentence — to an LLM that read the brain, executed against the right system, returned the answer, and wrote what was decided back. The work happens; the memory accumulates; the company gets quieter while doing more.
Each agent owns one domain. All eight share one brain and one master prompt. They read context, execute against the real systems, and write the patterns they learn back to the brain. Replace each agent's domain with yours — the pattern transfers.
Morning briefings, cross-domain synthesis, competitive scans, knowledge mining. Refreshes the world model weekly from real signals.
Ticket triage, WISMO replies, brand-voice drafts, pattern detection, escalation routing. AutoResearch loop tunes prompts on its own.
Weekly P&L, AR follow-ups, invoice reconciliation, treasury, variance alerts. Reads accounting, bank, and commerce in one pass.
Daily store reports, staffing signals, transfer flags, store comparisons. Pulls POS and traffic data, writes the daily brief.
Campaign analysis, segmentation, copy patterns, SEO opportunities, attribution. 1,114 historical campaigns indexed in the brain.
Sell-through, variant audits, markdown candidates, pricing, wholesale ops. Flags risk before it shows up in the P&L.
Absences, payroll prep, vacation balances, onboarding, expenses. Hard privacy stops on health, recruiting, and personal contexts.
Claude Code with all 44 MCP tools and the full brain. The founder operates every agent and every system from one prompt.
All eight share one memory. All eight write back what they learn. When the CS agent detects a pattern in refunds, the Merchandising agent sees it on Monday.
Most AI vendors quote numbers they can't defend. "10× productivity." "50× return." "Pays for itself in a week." Then you ask how they calculated it and the conversation gets vague. Here's exactly how we get to 18:1.
| CS Agent — triage, drafts, policy lookups | 20h |
| Founder time freed — briefings & decisions | 10h |
| Finance — P&L, AR, reconciliation | 8h |
| Merchandising — sell-through & sizes | 6h |
| Retail — daily reports & staffing | 5h |
| Marketing — campaigns & segments | 5h |
| Strategy — synthesis & knowledge | 5h |
| HR — admin & payroll prep | 3h |
| Total hours offloaded weekly | 62h |
€21/h loaded labor derived from personnel budget ÷ headcount ÷ ~2,000 working hours × role mix. Conservative. No revenue impact claimed — every number here is a cost avoided.
This is the operating manual of a company running on it now: tickets, books, stock, weekly learnings, shipped back into the playbook.
Not a course. Operating documentation, updated only when production teaches something worth shipping.
Production activity feed, honest ROI breakdown, real cost data, and a governed public snapshot of the swarm. No mockups, no vanity KPI theater, and no hidden spreadsheet math.
Most AI demos show one capability at a time. A system running for six months accumulates depth: corner cases, weird customer behaviors, recovery patterns, decisions that are too specific for a sales deck but too valuable to forget.
Self-evolving prompts. The CS agent mutates weak prompts and auto-promotes winners. 94.7% accuracy after 3 months.
Six domain agents plus blind peer review for high-stakes calls. Strategy, finance, CS, marketing, ops, merchandising, ~€1 per deliberation.
Cross-company pattern library with strict anonymization. New deployments start at 70%+ autonomy instead of zero.
Inbox → OCR → classify → reconcile → approve. Five minutes per invoice drops to <30 seconds.
Real-time CM3 per product across commerce, inventory, ads, analytics, and accounting. Every SKU gets a live margin number.
1,114 campaigns analyzed for subject line, body, CTA, and performance. Learned rules drive future drafts.
Contract-based execution for multi-step work. Every step has acceptance criteria and rollback rules.
Tracks visibility across ChatGPT, Perplexity, Claude, and AI Overviews. Mention rate improved from 35% to 60%.
The strategy hub has a capped corporate card. It buys approved tools and files receipts automatically.
The brain isn't a snapshot. It's a system that learns. Every decision, fix, refund, campaign result, and customer message gets captured. The longer it runs, the more your AI knows what your company would do — and the less you have to explain.
Generic AI. Asks for context every time. Most decisions still need a human in the loop.
Pattern-aware AI. Knows your refund policy, your tone, your vendors, your reasons. Most routine decisions execute themselves.
You can fork the playbook and the skills. The memory of your company has to be built — by your company.
The "AI for business" category is loud and mostly indistinguishable. It helps to be precise about what Compai is and isn't.
Compai is the operating manual for running a consumer SME on AI agents. Eight months in production. Free to read, fork, adapt. Every claim auditable.
The actual playbook, skills, prompts, templates, patterns, and lessons. Written inside an 8-figure consumer brand running on AI.
The brain primitive. The MCP server. The agents. The 12 advanced capabilities. The 32 production lessons. The honest open questions.
Executable procedures any AI client can read and run. Call them from a Claude window, an agent, or a script.
The actual prompts behind every agent: customer service, finance, marketing, retail, merchandising, people ops, strategy.
Brain skeleton, agent SOUL.md, factory.yml, custom instructions, and MCP integration starting points.
Anonymized operational patterns from a year in production. New deployments inherit, instead of starting from scratch.
What broke and what we did. Token leaks. Webhook downtime. Memory degradation. Each with a one-rule lesson others can use.
License: Free to use, attribution requested. Updated as production learnings are verified.
If you're 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.
45 chapters. 75 skills. 38 prompts. 32 production lessons. Free to use, attribution requested.