mm

A CLI that turns AI prompting from guesswork into engineering. Define what you want, delegate with precision, measure the results.

Install

npm i -g mm-kit   # installs both `mm` and `kaya` globally

Or from source:

git clone https://github.com/multimodeai/mm-cli.git
cd mm-cli
npm install && npm run build
npm link          # installs `mm` and `kaya` globally

Requires Node.js 22+ and a Claude API key or OAuth token:

export ANTHROPIC_API_KEY=sk-ant-api-...
# or
export CLAUDE_CODE_OAUTH_TOKEN=sk-ant-oat-...

The Flow

mm produces the documents that make AI coding sessions work. There are three phases:

Phase 1 — Define (you + mm)

mm spec new <feature>       →  specs/<feature>.md     # WHAT to build
mm intent init              →  INTENT.md              # HOW decisions get made
mm constraint <task>        →  constraints/<task>.md   # WHERE the boundaries are

Each command runs an interactive interview. Claude asks you questions, reads your codebase, searches the web for research, and produces a structured document.

If the output file already exists, mm enters edit mode — loads the existing document and asks what you want to refine. Use --fresh to start over.

Phase 2 — Build (your AI agent)

Take the spec and hand it to Claude Code, Cursor, or any AI coding tool:

"Execute specs/content-uniqueness.md. Reference INTENT.md for
decision authority and CLAUDE.md for project context."

The spec is precise enough for autonomous execution — acceptance criteria, task decomposition, constraint architecture, definition of done.

Phase 3 — Measure (mm again)

mm eval new <skill>                    # Build eval suite
mm eval run <skill>                    # Run with skill loaded
mm eval run <skill> --without-skill    # Run baseline
mm eval compare <skill>               # See the delta

A/B test your AI output with and without context engineering. Multi-axis 5-dimension scoring.

KAYA — review artifacts in the browser

mm ships with KAYA, a self-contained review surface for the documents it produces (and any HTML or Markdown artifact). Open a file, read it rendered on a dark themed page, toggle Annotate to attach comments to specific passages, and send the batch back to the agent. The agent/you conversation persists across rounds, so you see exactly what changed each iteration.

kaya path/to/plan.md         # open a review surface at a local URL
kaya path/to/report.html     # HTML artifacts work too
kaya export path/to/file     # write a self-contained offline .html (assets inlined)
kaya end path/to/file        # close the review and release the agent
kaya stop                    # stop the running server

KAYA runs entirely on your machine — a dependency-free Node server with vendored render assets, no network calls, no third-party host. Diagrams render as hand-drawn dark Mermaid, and the offline export inlines every asset into one file you can host anywhere.

Use it inside the spec loop to review and revise through the browser instead of the terminal:

mm spec new <feature> --kaya    # or: MM_REVIEW_ENGINE=kaya

All Commands

Command	What it does	Output
mm preflight	Print the 7 pre-prompting questions	stdout
mm diagnose	5-question AI workflow diagnostic	CONTEXT.md
mm diagnose --deep	12-question deep diagnostic + roadmap	DIAGNOSTIC.md
mm diagnose --health	Automated project health check — scores agent-readiness	HEALTH.md
mm rewrite	Rewrite vague requests into clear ones	stdout / REWRITE.md
mm context build	7-domain interview for business context	.claude/skills/business-context/SKILL.md
mm spec new [name]	Specification engineer (3-phase interview)	specs/<name>.md or SPEC.md
mm spec new [name] --type decompose	Break large changes into safe, ordered steps	specs/<name>.md
mm spec new [name] --type qa	QA specification with discovery + coverage math	specs/<name>.md
mm intent init	Intent & delegation framework	INTENT.md
mm constraint <task>	Constraint architecture (must/must-not/prefer/escalate)	constraints/<task>.md
mm eval new <skill>	Build eval suite via interview	evals/<skill>/eval.yaml
mm eval new <skill> --quick	Auto-generate eval from SKILL.md	evals/<skill>/eval.yaml
mm eval run <skill>	Execute eval suite	evals/<skill>/results/<ts>.json
mm eval compare <skill>	A/B comparison table	stdout
mm skill new <name>	Scaffold a new skill	.claude/skills/<name>/SKILL.md
mm skill list	List skills in current project	stdout
mm skill validate	Check skill structure	stdout
mm skill export --format cursor	Export skills to other IDEs	.cursorrules / .windsurfrules
mm harness verify [spec]	Verify codebase against spec	verify/<spec>/<ts>.json
mm harness audit	Lock-in audit (5 dimensions, /25)	HARNESS-AUDIT.md
mm harness audit --security	Security & resilience audit	SECURITY-AUDIT.md
mm harness route <task>	Task-to-harness routing advice	stdout
mm harness brief	Executive switching cost summary	HARNESS-BRIEF.md

Tools During Interviews

Commands that need codebase access (spec, eval, constraint, intent) give Claude tools to explore your project during the interview:

• read_file — Read any project file
• list_files — Find files by pattern
• list_directory — List directory contents
• search_files — Grep file contents
• git_info — Git log, diff, blame for repository context
• web_search — Search the web (DuckDuckGo, no API key needed)
• web_fetch — Fetch and read web pages

Claude reads your code before asking questions, and searches arxiv/docs when research is needed. Specs reference actual files, functions, and line numbers — not generic placeholders.

Global Flags

--model <model>    Override Claude model (default: claude-sonnet-4-20250514)
--dry-run          Print system prompt without calling API
--fresh            Ignore existing output file, start from scratch

How It Works

The interview engine sends prompt templates as Claude's system message. Claude drives the conversation — asks questions, reads your codebase, does research. The engine routes your answers back. When Claude produces the final artifact, it's auto-saved to disk.

┌─────────────────────────────┐
│  CLI Layer (Commander.js)   │
│  24 commands                │
└──────────┬──────────────────┘
           │
┌──────────▼──────────────────┐
│  Interview Engine           │
│  Multi-phase Claude         │
│  interviews → files         │
└──────────┬──────────────────┘
           │
┌──────────▼──────────────────┐
│  Claude Client              │
│  @anthropic-ai/sdk          │
│  Tool use + OAuth support   │
└─────────────────────────────┘

Background

Built on the insight that prompting split into 4 disciplines:

1. Prompt Craft — writing clear requests
2. Context Engineering — giving AI the right background
3. Intent Engineering — encoding decision-making rules
4. Specification Engineering — precise specs for autonomous execution

Skills + evals = measurable improvement in AI output quality.

License

FSL-1.1-MIT (Functional Source License) — free to use, modify, and distribute for any purpose except building a competing product. Each version converts to plain MIT two years after its release. See LICENSE.md for the full terms.

KAYA's vendored render libraries (DaisyUI, Tailwind, Mermaid) keep their own MIT licenses — see kaya-editor/vendor/THIRD-PARTY.md.