KeiSeiKit

A local-first substrate that installs next to any LLM-based assistant and keeps its memory, accountability, and nightly learning persistent. Rust core (48 crates, ≤ 2 MB each), TypeScript plugin glue. MIT.

Primary target: Claude Code. Also compatible with Cursor, Continue, Zed, Aider, Windsurf via 11 cross-tool bridges, and with any OpenCode / OpenClaude / open-source Claude-Code clone that speaks the MCP server protocol + standard hook events.

One-line install (Claude Code)

/plugin marketplace add KeiSei84/KeiSeiKit
/plugin install keisei@keisei-marketplace

Twelve agents, 43 skills, 12 hooks, nightly consolidation wired in 60 seconds. Install paths for MCP-only, Cursor, Continue, Zed, Aider, Docker, Nix — docs/INSTALL.md.

What it does (and where to verify)

DNA — persistent identity per agent run. Every invocation gets an 80-char deterministic string of the form:

<role>::<caps-bitmap>::<sha8-scope>::<sha8-body>-<hex8-nonce>

Example:

edit-local::NG-FW-FD-CP-CG-TG-ND-RF::5435F821::AC73A6A3-e9bf468d

• role — which capability bundle the agent got (edit-local / read- only / auditor / merger / …).
• caps-bitmap — which policy fragments composed its prompt (no-git- ops, files-whitelist, constructor-pattern, cargo-check-green, …).
• scope — 32-bit hash of the files-whitelist + denylist. Same files → same scope regardless of who asks.
• body — 32-bit hash of the task description. Two people asking the same thing → same body.
• nonce — random per spawn. Distinguishes reruns of identical tasks.

Same scope+body → same problem space. You can ask "has this type of task run before, how did it go" in one query, without embeddings, without a vector DB, without round-tripping through an LLM. The index is a plain 32-bit hash lookup against the SQLite ledger. kei-dna-index precedent --body <sha8> returns every past run with its status merged / failed / running — instantly. Code: _primitives/_rust/kei- shared/src/dna.rs (format SSoT), _primitives/_rust/kei-dna-index/** (queries), _primitives/_rust/kei-ledger/src/schema.rs (storage).

Git branch per work-unit. kei-fork create makes an atomic triplet: a branch, a worktree under _forks/<agent-id>/, and a ledger row. Agent writes in isolation, touches .DONE, kei-fork collect atomically merges + archives + updates ledger. On any step failure, everything rolls back. No more "the agent overwrote main". Code: _primitives/_rust/kei-fork/**.

Persistent memory on disk. Three SQLite files: ~/.claude/agents/ledger.sqlite (who-did-what), ~/.claude/memory/ sessions.sqlite (JSONL traces + embedding cache), ~/.claude/memory/ kei-memory.sqlite (cross-session patterns). All backupable with cp, portable with keisei mount, inspectable with kei-brain-view summary. Nothing in RAM-only; nothing in a cloud you don't control. Code: _primitives/_rust/kei-{ledger,memory}/**.

Prompts composed from fragments. Agent prompts aren't hand-written — they're assembled from _capabilities/<cat>/<slug>/text.md files at spawn time. Edit a fragment, every agent inherits the change. Single source of truth for prompt engineering. Code: _primitives/_rust/kei- agent-runtime/src/compose.rs.

Sleep — three phases of overnight consolidation. Your assistant keeps working while you're asleep, but on your rules, on your machine.

• Phase A — incubation. Anytime during the day:

  /sleep-on-it "write a 4-page literature review on …"
  

  adds a task to ~/.claude/memory/sleep-queue/. At 03:00 local, a scheduled agent picks up the queue (up to 5 tasks, ≤ 480 min total budget), works on each, checkpoints partial results every N minutes so a cut-short run still leaves a readable artefact.

• Phase B — REM. Same agent reads the day's JSONL traces under ~/.claude/memory/traces/, finds cross-session patterns ("you hit the same error three times this week", "these two tasks touched the same files"), and writes a morning report to ~/.claude/memory/ sleep-reports/YYYY-MM-DD.md. You git pull and read. Nothing auto-injects into your next session — decisions stay with you.

• Phase C — NREM, every 7 days. A conflict-scan checks the whole substrate for drift: rules that contradict, hooks that overlap, duplicate capability fragments. A refactor-engine then proposes a clean-up branch — in plan-only mode (markdown proposal) or plan + fork mode (actual git branch with patch applied). You pick during /sleep-setup.

Code + rule: docs/SLEEP-LAYER.md, hooks/ sleep-*.sh, _primitives/_rust/kei-conflict-scan/, _primitives/ _rust/kei-refactor-engine/.

Self-audit, silent-first. Three hooks — session-end-dump, milestone-commit-hook, error-spike-detector — feed a retrospective. Sessions 1–10 log without prompting. From session 11 on, if the same mistake reappears twice, /escalate-recurrence offers to codify it as rule + wiki entry + hook. Code: hooks/*.sh, skills/self-audit/, skills/escalate-recurrence/.

Deployment modes

Where	How	When you want it
Local laptop	/plugin install keisei	Daily dev / research / writing
USB stick	keisei mount /Volumes/usb/brain	Move between machines
iCloud / Dropbox	Same keisei mount on a cloud-synced folder	Auto-backup + multi-device
Cloud VPS	ssh user@host && ./install.sh && keisei serve	Always-on agent, tablet client, shared team brain
Docker	docker run -v ./brain:/home/kei/.claude keiseikit	Ephemeral CI runs, reproducible test envs
MCP-only	./install.sh --profile=mcp	Plug the memory layer into a different assistant, skip the plugin

Rust binaries are ≤ 2 MB each, statically linked where possible, no Python runtime. Any x86_64 / arm64 host with git + SQLite works. Bring your own API key (Anthropic / OpenAI / OpenRouter / local Ollama) via ~/.claude/secrets/.env — the kit never ships keys.

Why Rust + TypeScript (no Python)

Rust for the engine — 47 crates covering ledger, memory, DNA, fork, scheduler, watcher, assembler, CLI. The class of mistakes LLMs make most often — None vs empty array, missing .await, unhandled Result, mutex misuse — are compile-time errors in Rust. The LLM literally cannot ship them. The type system is the fence.

TypeScript only for the thin adapter that talks to Claude Code's Node-based plugin API. Three packages under _ts_packages/, each ≤ 300 LOC: marketplace manifest, installer helpers, plugin surface. The substrate itself never assumes Node is running.

Why not Python — it's transformer math, not taste. Writing Rust costs more tokens up-front (explicit types, explicit errors, explicit ownership) but debugging costs ≈ 0 — the compiler is the fence. Writing Python costs ~5 minutes (no types, no lifetimes, no compile step), then debugging loops for hours: each failed run produces new tokens the assistant must read, reason about, correct — often triggering the next failure, another iteration, another token burn. This isn't empirical. Python's tokenization (indentation-as-syntax, runtime dynamic dispatch, late-binding names) leaves each token with low semantic commitment, so the transformer must track more conditional paths forward. Rust pins each token to a compile-time shape: fewer open paths, fewer silent surprises, fewer debug loops. Net cost of a shipped Rust primitive: predictable. Net cost of a Python primitive: unbounded. RULE 0.2 enforces this with a hook that blocks python / python3 / uv run python without an explicit opt-in prefix. install.sh runs on a fresh host with zero Python installed — if Python slips into this repo, it's a bug.

Probably fits — it's a way of working with AI as much as a tool

You'll like this if you prefer:

• Seeing what your AI did, rather than a dashboard that hides the mistakes. The ledger shows every agent run, every fork, every rollback. Honesty by construction.
• Your own machine over someone else's cloud. Brain lives on your disk, backs up with cp, moves on a USB.
• Two install commands once, instead of a SaaS onboarding flow.
• One of: macOS, Linux, or Windows-via-WSL2. Native Windows is on the backlog.
• Owning your API keys — bring your own Anthropic / OpenAI / OpenRouter / local Ollama, never ours.
• Working across multiple AI assistants — Claude Code today, Cursor on the plane, an OpenCode fork next month; the substrate moves with you.
• A workflow that learns while you sleep rather than a tool you restart every morning.

If you'd rather have a hosted multi-tenant SaaS with a polished web UI and no CLI — this isn't that, and that's OK. Plenty of great tools in that space. KeiSeiKit is for the other camp.

What ships (run the command, get the number)

Thing	Count	Verify
Rust crates (workspace)	47	ls -d _primitives/_rust/*/ | wc -l
Tests green on cargo test --workspace	801	cd _primitives/_rust && cargo test --workspace
Agents	12	ls _agents/kei-*.md 2>/dev/null | wc -l
Skills	43	ls -d skills/*/ | wc -l
Hooks	12	ls hooks/*.sh | wc -l
Behavioural blocks	82	ls _blocks/*.md | wc -l
Capabilities	11	find _capabilities -name text.md | wc -l
Roles	7	ls _roles/*.toml | wc -l
Cross-tool bridges	11	ls _bridges/*.tmpl _bridges/*.mdc 2>/dev/null | wc -l

Counts regenerate from source via scripts/regen-counts.sh.

Architecture in one paragraph

One hub crate (kei-entity-store) owns SQLite migrations, FTS indexes, verb dispatch. Sibling store crates (chat, content, social, crossdomain, task, scheduler) are thin shims over the hub — star topology, zero sibling-to-sibling edges. DNA format is SSoT'd in kei-shared (grep-verify: grep -l "kei_shared::dna" _primitives/ _rust/*/src/*.rs). Constructor Pattern is enforced pre-commit: every file ≤ 200 LOC, every function ≤ 30 LOC. See docs/ARCHITECTURE.md.

Verify everything works (60 seconds)

git clone https://github.com/KeiSei84/KeiSeiKit && cd KeiSeiKit
cd _primitives/_rust && cargo test --workspace
# Expect: test result: ok. 801 passed; 0 failed
cd .. && tests/substrate_integration.sh && tests/hook_wiring_integration.sh
# Expect: two PASS lines

If both pass, the repo you're reading matches the repo I'm describing.

Compared to similar tools

	KeiSeiKit	LangChain	AutoGen	MCP alone
Local-first, no cloud required	✓	partial	✓	✓
Rust core, ≤ 2 MB binaries	✓	✗	✗	N/A
Per-run deterministic DNA	✓	✗	✗	✗
Git-fork work isolation	✓	✗	✗	✗
Three-phase nightly cycle	✓	✗	✗	✗
Capability-fragment prompts	✓	partial	partial	✗
Cross-tool bridges (Cursor, Zed, Aider, …)	11	✗	✗	✗
Works with OpenCode / OpenClaude clones	✓	✗	✗	✓
Enforced Constructor Pattern	✓	✗	✗	N/A

Sits alongside these tools, not a replacement. Uses MCP where MCP fits; LangChain stays useful for its retriever zoo.

Docs

PHILOSOPHY.md	The biological principles, in depth
INSTALL.md	All install paths, profiles, keisei CLI, hook controls
ARCHITECTURE.md	Build pipeline, bridges, meta-composer
REFERENCE.md	Every primitive, hook, skill with flags and exit codes
SLEEP-LAYER.md	Phase A / B / C nightly cycle + self-audit
TAXONOMY.md	The seven-facet vocabulary
SUBSTRATE-SCHEMA.md	Atom contract
SECURITY.md	Threat model + mitigations
USB-BRAIN-GUIDE.md	Portable brain — macOS / Linux / Windows
WHY.md	The full story of why this exists
CHANGELOG.md	What changed, version by version
PLUGIN.md	Anthropic plugin-format details

Author

Built by Denis Parfionovich at KeiLab (parfionovich@keilab.io) while running 4–8 parallel Claude Code terminals per day. Forks welcome.

MIT License. See LICENSE.