Cabinet governance
Separate AI roles propose, execute, critique, supervise, and leave decisions for each loop.
Japanese-primary · Multi-model AI · Governed runners
Naikaku brings planning, coding runners, sandbox execution, critique, supervision, and evidence-based iteration into one governed workbench.
Instead of handing work to one unchecked model, Naikaku separates planning, execution, critique, supervision, safety, memory, and scoring. Execution sits behind presets, readiness checks, logs, receipt review, and artifact audit.
Separate AI roles propose, execute, critique, supervise, and leave decisions for each loop.
Designed for OpenAI-compatible endpoints, OpenRouter, Anthropic, Aliyun DashScope/Qwen, Gemini, and local models.
Codex CLI, Claude Code, OpenHands, OpenClaw-style tools, Hammerspoon, and similar CLIs can be attached as governed runners.
Start locally, check provider aliases and the gateway, then attach runner presets only when you need them. Raw API keys should stay out of browser storage.
Get the macOS developer preview from GitHub Releases.
2Use npm ci and npm run dev to open the Workbench.
Run npm run gateway for local provider and runner routes.
Configure provider, endpoint, model, and API-key alias per role.
5Allow only the CLI runners you choose, with evidence required.
Try it, connect APIs, supervise a Mac runner, or contribute. Each entry point keeps the same language structure.
Run steps, verification commands, and checksums.
Bring-your-own keys and future hosted gateway.
Boundaries for CLIs and Mac automation.
Pick the entry point by intent.
Contributions, adapters, and verification.
Noncommercial source and commercial permission.
Japanese stays the primary language, while English and Chinese keep matching pages, matching routes, and matching license boundaries.
The same content structure is available.
The same content structure is available.
The same content structure is available.
It is an AI cabinet workbench where multiple roles plan, execute, critique, supervise, score, and preserve evidence.
Yes. Each role can use its own provider, endpoint, model, and API-key alias.
The public release is for noncommercial use. SaaS, client delivery, commercial modification, and paid integration require written permission from EMYSTI LLC.