What it is
KeiSeiKit is a coding-agent stack the lab built for itself and made public: a native terminal cockpit (file tree, embedded shell, live agent session panel) driven by an agent daemon that speaks to whichever model backend is configured — the lab's own Anthropic-native provider, or any OpenAI-compatible endpoint, local or remote — through the same provider trait. Agent identity, orchestration policy, and guard-rail hooks are described declaratively, not hard-coded per agent.
Everything through one cache
Underneath every agent capability call sits a single content-addressed cache — a SHA-256-keyed, SQLite-backed, TTL store that every atom-scale operation runs through. The same substrate gives the lab persistent cross-session memory and deterministic "has this exact request already been answered" lookups, instead of re-deriving the same result every session.
A sleep layer — agents improve between sessions, not only in one
Most agent tooling treats every session as a blank slate. KeiSeiKit runs a three-phase sleep cycle instead: an incubation phase that lets slower background work land, a REM-style morning report that surfaces what changed overnight, and a longer NREM pass that produces a rolling 7-day refactor plan from the accumulated session history. The agent's working state consolidates while it is idle, the same way the lab's nervous-system substrate (keineuron) reasons about stability over time — see the research page for that side of the story.
Why open
The lab uses these crates every day — 4 to 8 parallel agent sessions, daily. Keeping them public means the same code that powers the lab's own workflows is auditable, forkable, and improvable by anyone: "a tool first, not a product — if it solves your problem, fork it." A moat built on adoption, not secrecy.