Academic programs lose half their decisions in email threads and meeting notes—and new stakeholders can’t get up to speed without spelunking through inboxes. ProgramOS is an open spec for an AI program coordinator that fields questions across every stakeholder channel and captures every decision to a structured curriculum repository. Built at Gies to run the online MSBAi program, where it goes by the persona name K-ai.
NanoClaw is the engine. The work to make it a program coordinator lives outside the framework.
NanoClaw is the lightweight, container-isolated agent framework that handles channel registry, message routing, and Docker spawning. Adopters fork it directly from upstream.
Hand SPEC.md to a coding agent (Claude Code, Cursor, Codex) along with your fresh NanoClaw fork. The agent reads both, asks clarifying questions, and produces an adaptation matched to your program.
Most program-specific behavior lives in per-channel agent prompts (groups/<channel>/CLAUDE.md) and in the structure of your curriculum repository. Both stay readable, both stay editable.
One agent, six channels, one curriculum repository as source of truth
Cloned templates go stale. Specs stay useful.
NanoClaw upstream evolves. A fork-once template would freeze you at a point in time. With a spec, your agent reads the current NanoClaw and the spec together, producing an adaptation that matches both.
Your channels, sender allowlists, decision-authority rules, and curriculum structure are unique. A template would force you into someone else’s shape. The spec describes contracts, not code.
The reference deployment serves a real Master’s program with stakeholder PII, internal allowlists, and accreditation drafts. The spec is the part that generalizes—everything sensitive stays in adopter-owned private repos.