keisei

KeiSeiKit

A multi-LLM substrate for agentic coding. Same agent definition, any LLM backend — Claude Code, Grok, Antigravity (Gemini), GitHub Copilot, or Kimi. Pick your orchestrator with kei pick; agents spawn sub-agents on other LLMs via MCP spawn_agent; safety hooks enforce on every backend through a 3-tier model. Three-phase nightly sleep consolidates 30-session windows into morning markdown reports.

Apache 2.0 — explicit patent grant + retaliation clause.

Highlights

• 5 LLM CLIs unified. Claude Code (native hooks), Grok (port to ~/.grok/settings.json), Antigravity/Gemini, GitHub Copilot, Kimi. DNA-routed: each agent's manifest declares a provider; kei agent <name> resolves DNA → primary → claude fallback.
• Sub-agents on any backend. Agents call spawn_agent (built-in MCP tool in kei-mcp) to dispatch other agents to whichever LLM fits the task. Cross-CLI orchestration without lock-in — Grok can spawn critic@Claude, then ml-implementer@Gemini, all from one session.
• 3-tier policy enforcement. Claude + Grok TIER 1 (full native PreToolUse), Copilot TIER 2 (MCP-wrapped + --excluded-tools=shell), Agy + Kimi TIER 3 (advisory). no-github-push, safety-guard, destructive-guard, citation-verify, numeric-claims-guard surface on every backend that supports tool-call gating.
• Three-phase nightly sleep. Phase A (incubation — queued tasks via /sleep-on-it), Phase B (REM consolidation — analyzes last 30 sessions, writes morning markdown), Phase C (NREM deep-sleep, every 7 days — conflict scan + refactor proposals). Outputs are markdown; you decide what merges.
• Native token streaming. Each backend streams in its own print mode (claude -p, grok --print, agy --print, copilot --prompt); KeiSeiKit composes the agent prompt + task and passes through. No buffering layer.
• Persistent memory. SQLite ledger + content-addressable store, session-spanning context, cross-machine sync via memory-repo.
• Agent DNA. Deterministic variable-length identity per invocation: <role>::<caps>::<scope-sha8>::<body-sha8>-<nonce8>. Same task → same prefix → "did this run before?" via SQL, no embeddings.
• Constructor Pattern. Substrate, not framework. You compose; it doesn't dictate workflow. File >200 LOC → decompose. No mixins, no DI containers, no abstract factories in user code. (Backend dispatch via Box<dyn Trait> — e.g. kei-store::factory::build_store selecting between filesystem / s3 / github / forgejo / gitea memory backends — is canonical Rust and stays.)
• Self-maintaining. Every substrate edit cascades: registry updates, agent regeneration, DNA index refresh, keimd graph reindex. Auto-self-indexing via kei-registry SQLite.

By the numbers (v0.63.0)

109 Rust crates · 52 skills · 53 hooks · 37 agent manifests · 83 substrate blocks · 11 install profiles (minimal → full).

Platforms

• macOS (arm64 + x64) — fully supported, primary dev target.
• Linux (Ubuntu, Debian, Fedora, Arch — x64 + arm64) — fully supported.
• Windows — substrate itself is Bash-only, but the MCP server binary ships as kei-mcp-server-windows-x64.exe in every release. Three recommended paths:
  • WSL2 (recommended) — install Windows Subsystem for Linux, then run bootstrap.sh inside Ubuntu/Debian as normal. Full substrate works. bootstrap.sh auto-detects WSL2 and Git Bash; on bare Windows it prints a one-time WSL setup guide and copies the wsl --install command to your clipboard.
  • MCP-only — drop kei-mcp-server-windows-x64.exe into your Claude Desktop / VS Code MCP config to get spawn_agent + kei_bash/kei_edit/kei_write tools, without the full Bash-based substrate. Skills, hooks, and kei CLI not available in this mode.
  • Native PowerShell port — demand-driven. WSL gives 100% coverage today with 0 code duplication, so a native .ps1 substrate isn't built yet. If you want it, open an issue with a thumbs-up on the existing Windows-native tracker (or file one) — once demand is real, we'll build it. The MCP-server binary path already covers the common "just want spawn_agent in Claude Desktop" case.

Maturity matrix

The substrate ships as a layered set of components at different maturity levels. Read this before relying on any single piece for production work.