compound-operations-model

# The Compound Operations Model Open-source AI operations playbook for consumer SMEs, published by Compai. [](LICENSE) [](https://usecompai.com) [](https://usecompai.com/playbook) [](CONTRIBUTING.md) > **The operating system behind an 8-figure consumer brand running on AI for €631/month.** Read it. Fork it. Build your own. This repository documents the operating system pattern behind a real 8-figure consumer brand running AI agents in production: company Brain, MCP tools, domain agents, employee onboarding, health audits, shared memory, and controlled action queues. This is not a SaaS repo. It is an educational implementation portfolio: read it, fork it, adapt it, and keep humans in the approval loop where actions carry risk. ## Current Version **v2.2 - 10 June 2026** Latest update: hardening the brain. We ran a due-diligence audit on our own system and published the findings and fixes as [`chapters/10s-hardening-the-brain.md`](chapters/10s-hardening-the-brain.md): - Auth is live in `protect` mode: high-risk anonymous/invalid operations are blocked, bearer-key auth is active, and client migration is still ongoing. Current canonical status is tracked in the the company Brain at `knowledge/platform/tools/mcp-auth-status.md` (last verified 2026-07-03: 33.65% anonymous requests in the latest 2,000 auth-observe rows). Do not claim `enforce` until the live MCP service is actually running in `enforce`. - Resilience: git-versioned brain (15-min auto-commits + change ledger), dual encrypted backups, checksum-verified restores, batch rollback - The execution loop: backlog hygiene with anti-boomerang archiving, throttled generators, a daily triage digest hard-capped at 10 items - The sequencing rule: no new ingestion sources until the closed-loop rate is above 30% v2.1 (9 June 2026) added the operational compounding loop (`10r`): health audits, inbox sweeper, skill eval harness, skillify loop, shared memory contract, L3 action queues, and workflow mining. ## What Is Inside | Path | Contents | |---|---| | `chapters/` | Full public playbook source (56 chapters), including Brain v2, the operational loop, and brain hardening | | `kit/` | Starter implementation kit: init CLI, MCP server template, compliance scaffold, onboarding pack, SOUL templates, systemd templates, monitoring scripts | | `skills/` | **24 public skills** (anonymized from production) + [the full catalog](skills/CATALOG.md) of what the 352 are | | `pattern-library/` | Anonymized operational patterns and schema | | `case-study/` | Anonymized reference case study | ## Start Here 1. Read [`chapters/00-index.md`](chapters/00-index.md). 2. Read the architecture: [`chapters/03-architecture.md`](chapters/03-architecture.md). 3. Read the Brain chapters: [`chapters/10l-brain-v2-living-memory.md`](chapters/10l-brain-v2-living-memory.md) through [`chapters/10s-hardening-the-brain.md`](chapters/10s-hardening-the-brain.md). 4. Inspect the implementation kit: [`kit/README.md`](kit/README.md). 5. Use the patterns as a starting point, not as a blind install script. ## Operating Model The model has four layers: 1. **Brain** - structured company memory, source paths, world model, tasks, outputs, health. 2. **Tools** - MCP interfaces to business systems and operational data. 3. **Agents** - domain-specific agents with clear ownership, tools, and escalation rules. 4. **Compounding loop** - capture, triage, package, queue, audit, and write back. The key shift is bidirectionality. Agents do not only read the Brain. They write decisions, gotchas, workflow state, and completed work back into the shared memory. ## What Changed Since the Original v1 Repo The original public repo only contained three introductory chapters. This version publishes the full current playbook and public-safe starter artifacts. Important corrections: - Current brand: **Compai**. OperAI was the legacy/internal name. - Current repo: `usecompai.com/repo`. - The playbook is no longer just a multi-agent architecture essay. It now includes Brain v2, employee onboarding, setup 1-click, shared memory, health, and workflow mining. - ROI language now uses the audited 18:1 model from the public playbook, not older 31:1/50:1 marketing claims. ## What Not To Copy Blindly Do not copy a reference deployment's private data, credentials, Slack channels, Google Workspace accounts, HR records, customer records, or source-system tokens. Copy the contracts: - source paths on durable facts - owner and stale dates on docs - privacy hard-stops - human approval for risky actions - action ledger - health checks - explicit escalation rules ## Live Site - Playbook and dashboard: https://usecompai.com - Contact: hello@usecompai.com ## Community This is built in public, and it's meant to be copied. The best thing you can do is fork it, run it, and tell us what happened. - **Share what you built** — [open a showcase issue](../../issues/new?template=share-what-you-built.md). We keep a public list of real deployments. - **Improve the playbook** — fixes, clarifications, validated patterns, translations. See [CONTRIBUTING.md](CONTRIBUTING.md). - **Ask a question** — [open a question](../../issues/new?template=question.md) if a chapter left you stuck. - **Report a bug** — [bug report](../../issues/new?template=bug-report.md). Security issues go to [SECURITY.md](SECURITY.md), not public issues. Be useful and be kind: [CODE_OF_CONDUCT.md](CODE_OF_CONDUCT.md). ## License Creative Commons Attribution-NonCommercial-ShareAlike 4.0 International. See [`LICENSE`](LICENSE). Commercial usage or implementation help: hello@usecompai.com.