The PersonalOS Stack: 2026 Edition
Models, frameworks, and orchestration choices MASTEROS.ai ships by default — and the trade-offs behind each pick.
Every MASTEROS.ai PersonalOS ships on a default stack. Defaults exist for one reason: to ship fast and ship well. Every layer is documented and replaceable per client, but the defaults reflect what has worked across more than a hundred production deployments through early 2026.
Orchestration layer
Codex handles agent orchestration and the long-running execution loop. It owns tool calling, retries, human approval gates, and audit logging. We chose it for its programmability, its first-class evaluation primitives, and its predictable cost profile under load.
Reasoning and drafting models
- Long-context reasoning — frontier model from the principal's preferred provider, with zero-retention configured.
- High-volume drafting — a smaller, fast model fine-tuned on the principal's tone-of-voice corpus.
- Structured extraction — a deterministic model tuned for JSON output and schema adherence.
Knowledge layer
Postgres with pgvector is the default vector store, paired with a hybrid retriever that combines BM25, vector similarity, and a small reranker. Ingestion runs through a normalised document pipeline shared by every source — meetings, email, Drive, Notion, Slack — described in detail in our second-brain article.
Integration surface
Anything in the principal's existing stack — Gmail, Google Calendar, Google Drive, Calendly, Slack, Notion, Airtable, HubSpot, Salesforce, WhatsApp. We avoid greenfield workspaces. The PersonalOS lives where the work already happens.
Observability and evaluation
Every agent invocation is logged with inputs, retrieved context, model output, and downstream actions. An evaluation harness runs nightly against a golden set per agent. Regressions block deploy. This is the single most under-invested layer in DIY builds, and the reason most of them silently degrade.
Hosting and data residency
Everything is deployed inside the principal's own cloud accounts — typically a dedicated workspace in their Google or Microsoft tenant plus a private cloud project for the orchestration runtime. MASTEROS.ai never stores client data.
What we deliberately do not ship
- Greenfield SaaS dashboards. The principal already has too many tabs.
- Chat-only interfaces as the primary surface. Conversation is one channel, not the product.
- No-code workflow builders. They look flexible and end up brittle at scale.
Frequently asked
Can you use my preferred model provider?
Yes. Provider choice is configurable per deployment. We benchmark on your real workload and recommend, but the principal decides.
How often does the stack change?
Quarterly review. Production deployments only get upgraded behind a passing evaluation harness.
Is the stack open source?
The orchestration patterns, evaluation harness, and document pipeline are documented and portable. Client-specific code is owned outright by the client.