Okhai's Foundation builds a stricter stack for AI-era software

Why it matters

Foundation shows how independent builders are responding to AI coding: less faith in prompts, more investment in contracts, performance gates and agent operating rules.

Okhai. nmxmxh, an Abuja-based developer who describes himself as a "lazy ambitious developer, systems thinker, giant problem solver," has put Foundation in public as a work-in-progress full-stack application substrate for teams building high-performance, event-driven systems with AI coding agents in the loop.

The project is small by public traction metrics and large by ambition. GitHub listed the nmxmxh/foundation repository as public with 55 commits, no forks and a single main branch when checked on July 4, 2026. The README identifies the project as "Ovasabi Foundation," version 0.0.1, and describes it as an integrated toolkit that combines platform modules, scaffolds, enforcement checks and documentation for production systems.

That framing matters because Foundation is not being presented as another code-generation wrapper. Okhai is trying to make the underlying application stack more deterministic before the agent starts editing it. The repo's documentation tells developers and AI assistants to read a defined sequence of architecture documents, identify invariants before changes, and leave durable evidence through tests, benchmarks, static checks or migration proof. In practice, Foundation treats AI coding as an operations problem: agents can help, but only inside a system that makes boundaries, contracts and regressions explicit.

Okhai's public biography is thin, which is often the case with independent infrastructure builders before a project has users, customers or investors. His GitHub profile lists 70 repositories, 19 followers and a location in Abuja, Nigeria. The pinned work around Foundation gives more useful context than a resume would: master-ovasabi describes INOS as a modular backend built with a frontend mindset, while inos_v1 describes a zero-copy distributed runtime using SharedArrayBuffer, a Go kernel and Rust WASM compute modules. Foundation looks like the attempt to consolidate those instincts into a reusable stack.

The bet: make the contract the program

Foundation's core claim is that modern software wastes too much work translating the same state across databases, backends, network payloads, frontend stores and browser storage. The philosophy document, dated June 30, 2026, calls this the "software deficit" and argues for a single immutable state event as the common unit across the stack.

The implementation described in the README spans server-kit in Go, runtime-transport in TypeScript, runtime-sdk in Rust/WASM, runtime-native for a Tauri/Rust shell bridge, frontend-kit, ui-minimal, config-contracts and templates. The data layer described by the README uses PostgreSQL as durable truth, Redis for coordination, Protocol Buffers for contracts and Cap'n Proto for zero-copy boundaries.

The most concrete part of the project is its performance model. Foundation's README lists seven performance planes, from same-process direct dispatch at 10 to 30 nanoseconds per operation to JSON compatibility around 30 microseconds per operation, and says the fastest lane should not pay the cost of the compatibility lane. Those figures are project-stated benchmarks, not independent measurements. The important engineering choice is the enforcement posture: performance claims are supposed to be measured and guarded, rather than left as README copy.

Agent rules are part of the product

Foundation's AI angle sits in AGENTS.md, which calls itself a universal instruction file for AI coding assistants including Claude, Codex, Cursor, Copilot and Windsurf. The file tells agents to read Foundation's glossary, quick start, architecture contract, operating contract, practice controls and AI threat model before architecture-sensitive edits.

The repo's required definition of done for agent-authored changes is unusually operational. It asks the agent to state whether a public contract changed, identify the invariant that must still hold, leave evidence, preserve or document a fallback path, name the scope boundary touched, add or update a regression guard, and update docs or explain why no documentation changed.

That is the more interesting founder choice. Most AI developer tools chase distribution at the editor, pull request or CI layer. RuntimeWire has tracked adjacent projects such as Daniel Lyons' Treedocs, which makes repo maps fail when they go stale, and AISlop, which puts AI-generated code smells into CI. Foundation goes lower in the stack. Okhai is trying to shape the app architecture so agent mistakes become easier to constrain before they land.

The quick start makes the same point for human developers. Before editing, contributors are told to answer which layer owns a file, which invariant could break, which check or test would catch a recurrence, and which fallback remains when the fast path or external dependency fails. That reads like internal platform discipline packaged into a public repo.

The license is part of the business signal

Foundation is public source, but the license is not a simple permissive license. The OVASABI Foundation Community License Agreement v1.1 says Apache 2.0 terms are incorporated only within the permitted scope, with restrictions controlled by the custom agreement.

The license permits non-commercial use, open-source development under an OSI-approved license for derivative work, and commercial use only while an organization stays below a revenue threshold. That threshold is defined as more than $1 million in gross revenue over the most recent 12 months or more than $1 million in total financing raised. Above that line, commercial users need a separate commercial license from OVASABI STUDIOS LTD.

The AI clause is sharper. The license prohibits training use of the software, including fine-tuning, alignment, evaluation for development purposes, distillation and use as seed material for synthetic training data. It expressly permits inference use for developing, maintaining, reviewing, testing, documenting or operating software, including AI coding assistants and autonomous coding agents, provided the user does not knowingly grant the AI provider rights to retain or use the software for training.

That puts Foundation in a growing bucket of source-available developer infrastructure whose authors want AI-assisted usage while withholding model-training rights. Okhai is allowing agents to operate on the repo. He is drawing the line at turning the repo into training fuel.

What is still unproven

Foundation has no disclosed funding, revenue, customer list, package download data or company launch around it. The repo does not establish a venture-backed startup, a commercial product launch or a production customer base. It establishes a serious independent software project with a public architecture thesis and recent commit activity.

The commit history shows active work through July 3, 2026, including changes to Coolify-owned routing, a refresh command for the Ovasabi CLI, seed drift management, PostgreSQL and Redis Docker configuration, generated protobuf types, FileStore and ObjectStore tests, and Rust performance-unit documentation. That is enough to treat Foundation as a live project. It is not enough to treat the performance model as proven at customer scale.

The unanswered adoption question is simple: whether teams will accept this much structure at the start of a project. Foundation asks developers to buy into contracts, tenant isolation, lifecycle events, bounded work, static checks, performance lanes and agent handoff requirements before the product code has much room to move. That discipline is exactly what large systems need after they hurt. The harder sell is convincing small teams to accept it before the pain arrives.

Still, the timing is right for this kind of experiment. As coding agents write more application code, the valuable founder work shifts toward the rails those agents run on: schemas, invariants, tests, boundaries and rollback paths. Okhai's Foundation is early, narrow in public adoption and heavy on doctrine. It is also one of the clearer examples of an independent developer treating AI software development as a systems design problem rather than a prompt design problem.