Kamaʻāina

An AI Development Kit (ADK) for local-first AI on consumer hardware.

Skills-based agentic development — the Claude Code style of packaged, instruction-driven AI tooling — quietly assumes frontier infrastructure: enormous context windows, provider-side optimization, cheap parallel subagents. Open-weight models on consumer and prosumer hardware get none of that, and won't for the foreseeable future. Kamaʻāina's bet is that this is a design problem, not a waiting problem: like game development in the 70s–80s, the limits are the design space. This kit collects the principles, patterns, and tools for building agentic skills that fit — small contexts budgeted like scarce RAM, external memory instead of long sessions, deterministic tools wherever a model isn't actually needed.

Built with frontier assistance today; usable entirely within a local stack the moment you download it. (A frontier model can use this ADK too — it's just an over-provisioned consumer of the same artifacts.)

The two artifact classes

Everything here is exactly one of:

Class What Where Audience
1 Human-readable documentation — research, theory, best practices, design rationale docs/, design/ Humans (and frontier models helping them)
2 Downloadable tools — SKILL.md + manifest.yaml, runnable at a declared tier with no frontier dependency tools/ Local AI stacks

Tiers at a glance

Every tool declares the minimum hardware/model floor it needs. Full definitions in docs/tiers.md; the standing rule is design for the lowest tier that can do the job.

Tier Class Model Usable context
1 Laptop ~7–9B 8k–16k
2 Prosumer ~14–32B ~32k
3 Workstation (DGX Spark-class) 70B+ / large MoE ~128k

Principles

  1. Local-first — no frontier calls at use-time, network: none by default.
  2. Not-AI when possible — a calculator beats an LLM at arithmetic; deterministic tools first, model tokens reserved for judgment. Teach to fish.
  3. Transparency — every tool's manifest declares everything it touches; you can always be your own mechanic.

The full argument: docs/philosophy.md. Kamaʻāina is also deliberately stack-agnostic — it never encodes how your machine serves models; the bridge is always an OpenAI-compatible endpoint and a model name (the boundary).

Start here

  1. Philosophy — why this exists and the three principles.
  2. Tiers — the hardware floors everything is designed against.
  3. The same skill, built two ways — a worked example of a frontier-habit build failing at the tiers, and the Kamaʻāina build that doesn't. The fastest way to feel what this ADK is for.
  4. Context engineering patterns — the pattern catalog, with real worked examples and anti-examples.
  5. Loea design — the skill-creator that enforces all of the above: manifest schema · permission model · context budgets.
  6. Roadmap · Contributing.

About the name

Kamaʻāina — "child of the land," a longtime local — names the goal: AI tooling that belongs to the machine it runs on. The project's names honor the author's upbringing in Hawaiʻi and follow proper Hawaiian orthography in prose (ASCII in paths); the full note and glossary are in docs/naming.md.

Status

Early. Foundation docs and the first tool (Waihona, the context-base) are in; Loea v0 is next. Follow the roadmap and issues.

License

MIT.