mewkit

CLI for MeowKit — scaffold, migrate, upgrade, and manage your AI-agent toolkit. Requires Node.js ≥ 24.

Install

npx mewkit init                  # provider picker (Claude Code / Codex / Cursor), then scaffold
npx mewkit init --target codex   # skip the picker — scaffold a Codex-native toolkit
npx mewkit <command>             # runtime commands (see below)

Commands

| Command | Description | | --- | --- | | mewkit init | Scaffold or update a project. --target <provider>, --profile, --dry-run, --force, --beta | | mewkit upgrade | Update to the latest release; auto-refreshes installed provider toolkits. --check, --beta, --list | | mewkit migrate <tool> | Export the .claude/ setup to Codex or Cursor. --dry-run, --force, --global, … | | mewkit pack | Manage install packs: list / add <pack> / remove <pack> | | mewkit providers | Show the supported-tool matrix. --json, <provider> | | mewkit memory | View/manage .meowkit/memory: --show, --stats, --clear, capture | | mewkit doctor | Diagnose environment (Node, Python, Git, .meowkit/ / .claude/ layout) | | mewkit validate | Verify .claude/ structure integrity. --packs | | mewkit setup | Guided configuration (Python venv, MCP, .env, .gitignore) | | mewkit budget | Token usage + cost. --monthly; budget context for per-profile size | | mewkit capabilities | Resolve installed skills/commands for an intent: list, explain, resolve, projections | | mewkit visual-plan | Structured visual plan: validate / approve / edit / view + feedback loop | | mewkit task | new / list — structured task files | | mewkit status | Print version, channel, and config |

Init & targets

On a fresh install, mewkit init opens a provider multiselect — Claude Code (preselected), Codex, Cursor — and provisions each picked toolkit (Claude Code installs .claude/ from the latest GitHub release). --target skips the picker.

• mewkit init --target codex — create a Codex-only project by copying the authored Codex bundle (AGENTS.md, .codex/{config.toml,agents,hooks.json,hooks}, .agents/skills/). No .claude/, no conversion.
• mewkit init --target cursor — unpack .claude/, then export to Cursor.
• mewkit init --profile <name> — install a right-sized subset (see Profiles).

Migrate

mewkit migrate <tool> converts a Claude Code setup (agents, commands, skills, config, rules, hooks) into the target's native surfaces. Supported targets: codex, cursor. Idempotent — re-running over an unchanged source is a no-op, user-edited targets are kept unless --force, and every run previews with --dry-run.

npx mewkit migrate codex --dry-run     # plan + reference report, writes nothing
npx mewkit migrate codex               # migrate into the current project
npx mewkit migrate codex --include-mcp # also convert .mcp.json → [mcp_servers]

What migrates where for Codex:

| Source | Codex target | | --- | --- | | .claude/agents/*.md | .codex/agents/*.toml | | .claude/commands/**/*.md, .claude/skills/** | .agents/skills/*/SKILL.md | | CLAUDE.md, .claude/rules/*.md | AGENTS.md (merged, 32 KiB budget) | | .claude/hooks/*.cjs | .codex/hooks/ + hooks.json | | .mcp.json (opt-in) | config.toml [mcp_servers] (--include-mcp) |

Markdown references are rewritten with a fence-aware classifier: inline refs are repointed, runnable fenced commands are rewritten only when the referenced asset migrates in the same run (otherwise preserved and listed in the dry-run report), and citations stay verbatim. Generated files carry no toolkit branding.

Memory

Project memory lives in the runtime-neutral .meowkit/ state root (memory/, telemetry/, state/, cache/), so the same state works across providers. Every structured write goes through one validated contract (schema → secret scrub → injection scan → lock → atomic replace). mewkit migrate safely relocates any legacy .claude/memory/, and mewkit memory capture routes a ##pattern / ##decision / ##note prompt into the right store.

Profiles & Packs

init --profile <name> installs a right-sized subset. Every profile ships the safety base (gate/privacy hooks, security rules, settings, statusline, core commands).

| Profile | Contents | | --- | --- | | core | base + lifecycle essentials + memory + utility | | developer | core + testing + git + docs | | product | developer + product / autobuild | | atlassian | developer + Jira / Confluence | | security | core + security audit | | research | core + research / brainstorming | | full | everything (default) |

Add or drop domains with mewkit pack add|remove (remove preserves shared, base-covered, and user-edited files). mewkit budget context estimates each profile's loadable context size; mewkit validate --packs checks manifest coherence and the safety invariant.

Visual Plan

Structured visual review for UI-bearing plans. A plan's visual-plan/plan.json is a coverage ledger + canvas (lanes, artboards, semantic-HTML wireframes, connectors, annotations); the CLI validates and gates it, and a transient 127.0.0.1 React studio renders and edits it.

npx mewkit visual-plan validate <plan-dir> [--json]      # schema + coverage + refs + safe-HTML + hashes
npx mewkit visual-plan approve <plan-dir> --revision <n> # single writer of review.status (Gate 1)
npx mewkit visual-plan view <plan-dir>                   # read-only studio
npx mewkit visual-plan edit <plan-dir>                   # editable (single-editor lock)

Security: the studio binds 127.0.0.1 only, is Host-header guarded against DNS rebinding, blocks path traversal, and enforces a strict CSP (script-src 'self'). Wireframe HTML is DOMPurify-sanitized at save and render. The loopback primitives in src/local-web/ are ported in part from patoles/agent-flow under Apache-2.0 — see NOTICE.

Related

• GitHub · Docs

License

MIT