Kaizen
Architecture-driven modernization. Audit trail ships by default. Decompose legacy codebases into editable ADRs, recompose to modern stacks — compliance falls out of doing the architecture right, not bolted on after.
Install
| Method | Command |
|---|
| winget (Windows) | winget install Kaizen3C.KaizenCLI |
| npm | npm install -g kaizen-3c-cli |
| Homebrew (macOS/Linux) | brew tap Kaizen-3C/tap && brew install kaizen-cli |
| pipx (Python, recommended) | pipx install kaizen-3c-cli |
| uv | uv tool install kaizen-3c-cli |
| pip | pip install kaizen-3c-cli |
winget, npm, and Homebrew install a standalone binary (no Python required). pipx/uv/pip install the full package including [web] and [mcp] extras.
export ANTHROPIC_API_KEY=... # or OPENAI_API_KEY, or a local Ollama host
kaizen memsafe-roadmap ./my-c-lib --output roadmap.md
Does it work? — measured outcomes
| Case study | One-shot LLM | Kaizen (plain ADR) | Kaizen (+ domain schema) |
|---|
inih — C → Rust, 522 LOC (cargo check) | 6 errors | 1 error | 0 errors |
Nancy — NancyContext.cs .NET Fx → .NET 8, 148 LOC (dotnet build) | 14 errors | 0 errors | 0 errors |
Methodology: three-arm ablation (one-shot control, plain ADR pipeline, ADR + domain schema) on real OSS repos, measured with cargo check and dotnet build respectively. The full case-study write-ups — exact commands, prompts and raw outputs — are not published in this repository yet.
What it isn't
Kaizen is not a code-generation tool. The ADR is the product; the LLM is the tool that produces it. An ADR is an editable, auditable record of an architectural decision — the thing a compliance officer signs off on, the thing a reviewer pushes back on, the thing that survives when the model changes next quarter. Recompose is the CI check that the ADR actually predicts working code.
The two wedges
kaizen memsafe-roadmap — C/C++ → Rust. Produces a CISA-format memory-safety roadmap + per-module ADR stubs. Optionally recomposes to a Rust port.kaizen migrate-plan --from X --to Y — framework modernization (9 pairs: AngularJS → Angular/React, jQuery → React, .NET Framework → .NET 8/9, Python 2 → 3, Java 8 → 17/21, Spring 4 → Spring Boot 3).
Full CLI reference: cli/README.md.
License
Kaizen is dual-licensed:
- Apache-2.0 for the CLI, pipeline, and provider adapters —
cli/, cli/pipeline/, agents/src/providers/. Free forever. Installable via pip install kaizen-3c-cli.
- Kaizen Enterprise Commercial for the enterprise wrapper — multi-tenancy, RBAC, SSO, MFA, audit-log export, approval workflows, cost attribution UI, budget caps. Lives under
interface/ and is priced separately.
The open-core boundary is explicit and enumerated: everything under cli/ is Apache-2.0. No pipeline capability is paywalled — the commercial tier wraps infrastructure, not pipeline.
Documentation
Contributing
Contributions are welcome. See CONTRIBUTING.md. All Apache-2.0 inbound; Commercial-tier work happens in the private dev repo, not via public PRs.
Security
To report a vulnerability, follow SECURITY.md. Do not open a public issue for security-related findings.
For contributors — internals
Architecture
Kaizen uses a four-tier architecture with a multi-agent denoising convergence loop. The CLI + pipeline (top-left in the diagram) is the Apache-2.0 surface; the ASP.NET + React UI is the Commercial wrapper.
┌─────────────────────────────────────────────────────┐
│ React / TypeScript UI (interface/ui/) │ ← Commercial
│ Vite, Zustand, TanStack Query, Monaco, Recharts │
└────────────────────┬────────────────────────────────┘
│ REST / HTTP
┌────────────────────▼────────────────────────────────┐
│ C# / ASP.NET Core 9 API (interface/) │ ← Commercial
│ PostgreSQL, Redis, JWT + MFA │
└────────────────────┬────────────────────────────────┘
│ gRPC (deferred — see ADR-0061)
┌────────────────────▼────────────────────────────────┐
│ Rust Orchestrator (core/) │ ← Apache (deferred modules)
│ Concurrent decomposition engine, convergence gate │
└────────────────────┬────────────────────────────────┘
│ in-process / gRPC
┌────────────────────▼────────────────────────────────┐
│ Python Agent Services (cli/, agents/) │ ← Apache — the shipped product
│ Researcher, Red Team, Draft, Write, Evaluator │
│ LLM provider abstraction (Anthropic, OpenAI, │
│ Ollama, LiteLLM) │
└─────────────────────────────────────────────────────┘
The five agents (Researcher, Red Team, Draft, Write, Evaluator) iterate over the decomposition output. Each round produces grounded confidence scores (test pass rate, static analysis, coverage); the orchestrator gates completion via a Thompson-sampling convergence decision (see ADR-0052).
Developer setup
For CLI-only development, pip install -e . at the repo root is enough. See the CLI guide for command-level walkthroughs.