@kairox_ai/tastekit-cli

Pre-1.0 CLI for TasteKit — compile human taste into portable AI agent artifacts.

One interview. 32+ compatible runtimes.

TasteKit is still pre-1.0. Treat this package as a serious release candidate, not a battle-tested stable runtime contract.

Install

# Use directly (no install)
npx @kairox_ai/tastekit-cli init

# Or install globally
npm install -g @kairox_ai/tastekit-cli

Quick Start

# 1. Create a workspace (pick domain + depth + LLM provider)
tastekit init

# 2. Run the LLM-driven onboarding interview
tastekit onboard

# 3. Compile your taste into artifacts
tastekit compile

# 4. Export to your runtime
tastekit export --target claude-code
tastekit export --target openclaw
tastekit export --target agents-md

Commands

| Command | What it does | |---------|-------------| | tastekit init | Create workspace. Pick domain, depth, LLM provider | | tastekit onboard | LLM-driven adaptive interview (voice mode optional) | | tastekit compile | Generate artifacts from session (supports --resume) | | tastekit export | Export to claude-code, openclaw, manus, autopilots, agents-md, agent-file | | tastekit skills | list, lint, graph, pack, report, inspect, history, rollback | | tastekit drift | detect, review, accept, reject — behavioral drift monitoring | | tastekit mcp | add, list, inspect, bind — MCP tool management | | tastekit trust | init, pin-mcp, pin-skill-source, audit | | tastekit eval | run, replay — evaluation packs | | tastekit import | Import from SOUL.md or Agent File (.af) | | tastekit completion | Shell completions for bash, zsh, fish |

Runtime Compatibility

TasteKit's direct exports cover Claude Code, OpenClaw, Manus, Autopilots, AGENTS.md, and Agent File. The "32+ compatible runtimes" claim comes from combining those direct adapters with the broader AGENTS.md ecosystem: any agent or tool that reads AGENTS.md can consume the generated operating contract even when it does not have a dedicated TasteKit adapter.

Domains

| Domain | Dimensions | Description | |--------|-----------|-------------| | General Agent | 6-18 | Cross-functional agent configuration | | Development Agent | 9-28 | Engineering philosophy, code quality, testing, commits | | Content Agent | 7-23 | Brand voice, editorial standards, publishing | | Research Agent | 7-25 | Methodology, source evaluation, synthesis | | Sales Agent | 7-25 | Sales philosophy, objection handling, pipeline | | Support Agent | 7-25 | Escalation criteria, empathy, SLA targets |

Each domain supports three depth tiers: Quick, Guided, and Full Taste Composition. Full is coverage-driven rather than time-boxed.

Output

After compile, your .tastekit/ directory contains:

.tastekit/
├── constitution.v1.json           # Principles, tone, tradeoffs, taboos
├── guardrails.v1.yaml             # Permissions, approvals, rate limits
├── memory.v1.yaml                 # Retention policy, salience rules
├── skills/                        # SKILL.md files
├── playbooks/                     # Multi-step workflows
└── derivation.v1.yaml             # Compilation state

After export --target claude-code:

export/
├── CLAUDE.md                      # System instructions
├── .claude/settings.local.json    # Hooks + permissions + defaultMode
├── .tastekit/hooks/*.sh           # 5 lifecycle hooks
└── skills/                        # Agent Skills-compatible directory

Part of TasteKit

• Core library: @kairox_ai/tastekit-core
• Adapters: @kairox_ai/tastekit-adapters
• Validator: @kairox_ai/tastekit-validator
• Voice: @kairox_ai/tastekit-voice

License

MIT