Skip to content
Mongez — production TypeScript packages Mongez

Agent integrations

You’re a developer who wants their AI coding agent to read the same project conventions as the rest of the team and pick up the skills bundled inside the npm packages you depend on. This page is the per-IDE cookbook. Pick your agent below, copy-paste the snippet, ship.

Every section follows the same three-step pattern — what differs is which files agent-kit writes, which --target flag to pass, and what reload step the agent needs.

The shared baseline

Before any per-IDE step, every project does the same bootstrap:

Terminal window
npx agent-kit init

That writes a starter AGENTS.md at the project root (only if one doesn’t already exist) and derives CLAUDE.md, .gemini/GEMINI.md, .github/copilot-instructions.md, and CONVENTIONS.md from it. From here, the per-agent sections below differ only in which --target you pass to sync and which directory the skills land in.

Once you’ve picked your target(s), wire sync into postinstall so every future yarn install / npm install keeps everything current:

{
  "scripts": {
    "postinstall": "agent-kit sync"
  }
}

Using agent-kit with Claude Code

Claude Code is Anthropic’s terminal/IDE coding agent. It reads CLAUDE.md for project conventions and .claude/skills/ for installable skills. Live skill reload — the only agent that picks up new skill files within the same session without a restart.

Terminal window
# One-time setup
npx agent-kit init
npx agent-kit sync --target claude
{
  "scripts": { "postinstall": "agent-kit sync" },
  "agentKit": { "targets": ["claude"] }
}

agent-kit creates: CLAUDE.md (derived from AGENTS.md) + .claude/skills/<pkg-slug>-<skill-name>/SKILL.md for every skill bundled in your dependencies.

Reload quirk: if .claude/skills/ didn’t exist when your Claude Code session started, restart Claude once after the first sync so it discovers the new directory. After that, edits to existing skill files are picked up on the next prompt — no restart.

Using agent-kit with Cursor

Cursor is a VSCode-based AI editor. Reads root AGENTS.md natively (no derivation needed) and .cursor/skills/ for skills.

Terminal window
npx agent-kit sync --target cursor
{
  "scripts": { "postinstall": "agent-kit sync" },
  "agentKit": { "targets": ["cursor"] }
}

agent-kit creates: .cursor/skills/<pkg-slug>-<skill-name>/SKILL.md for every bundled skill. No CLAUDE.md-style derived file is generated — Cursor reads AGENTS.md directly.

Reload quirk: Cursor reads its .cursor/skills/ directory on session start. After running sync, close and reopen the project (or Cmd/Ctrl+Shift+P → “Reload Window”) so the new skills are picked up.

Using agent-kit with Codex

Codex is OpenAI’s coding agent (CLI + IDE flavors). Reads root AGENTS.md natively and .codex/skills/ for skills.

Terminal window
npx agent-kit sync --target codex
{
  "scripts": { "postinstall": "agent-kit sync" },
  "agentKit": { "targets": ["codex"] }
}

agent-kit creates: .codex/skills/<pkg-slug>-<skill-name>/SKILL.md. No derived AGENTS.md alternative needed.

Reload quirk: Codex picks up skills on the next session. CLI users get fresh reads on every codex invocation; IDE users should reload the window.

Using agent-kit with Kiro

Kiro is Amazon’s AI-native IDE. Reads root AGENTS.md natively and .kiro/skills/ for skills.

Terminal window
npx agent-kit sync --target kiro
{
  "scripts": { "postinstall": "agent-kit sync" },
  "agentKit": { "targets": ["kiro"] }
}

agent-kit creates: .kiro/skills/<pkg-slug>-<skill-name>/SKILL.md. No derived file.

Reload quirk: Reload Kiro (window reload) after the first sync. Subsequent skill edits typically need a fresh session too.

Using agent-kit with Gemini CLI

Gemini CLI is Google’s terminal-based coding agent. Reads .gemini/GEMINI.md (it does not read root AGENTS.md natively).

Terminal window
npx agent-kit sync --derive-only

agent-kit creates: .gemini/GEMINI.md derived from AGENTS.md (regenerated on every sync, so the two never drift).

No --target gemini for skills — Gemini CLI doesn’t currently have a published skills directory convention, so agent-kit doesn’t write one. Only the derived GEMINI.md is generated for this agent. If you also use Claude or Cursor on the same project, run sync --target claude,cursor so the derive step + the skills targets you do care about run in one shot:

{
  "scripts": { "postinstall": "agent-kit sync" },
  "agentKit": { "targets": ["claude", "cursor"] }
}

The Gemini derive runs on every sync regardless of targets — the targets array gates only the skills export. So this agentKit block keeps GEMINI.md fresh AND mirrors skills into Claude + Cursor.

Reload quirk: Gemini CLI re-reads its instructions on every invocation, so there’s no “reload window” step.

Using agent-kit with GitHub Copilot

GitHub Copilot is the VSCode / Visual Studio AI completion + chat agent. Reads .github/copilot-instructions.md for project conventions and .github/skills/ for skills.

Terminal window
npx agent-kit sync --target copilot
{
  "scripts": { "postinstall": "agent-kit sync" },
  "agentKit": { "targets": ["copilot"] }
}

agent-kit creates: .github/copilot-instructions.md (derived from AGENTS.md) + .github/skills/<pkg-slug>-<skill-name>/SKILL.md.

Reload quirk: Cmd/Ctrl+Shift+P → “Developer: Reload Window” after the first sync so VSCode re-scans .github/skills/. Copilot’s instructions file is re-read on every prompt, so no reload needed for AGENTS.md edits.

Using agent-kit with Aider

Aider is a terminal-based AI pair programmer that doesn’t follow the skills/ convention. It reads a CONVENTIONS.md file (loaded via /read or auto-loaded via .aider.conf.yml).

Terminal window
npx agent-kit sync --derive-only

agent-kit creates: CONVENTIONS.md derived from AGENTS.md.

To auto-load it in every Aider session, drop this in .aider.conf.yml at the project root:

read:
  - CONVENTIONS.md

No --target aider for skills — Aider doesn’t currently have a skills directory convention. Only the derived CONVENTIONS.md is generated for this agent.

Reload quirk: Aider re-reads CONVENTIONS.md on every session start.

Using agent-kit with Antigravity

Antigravity is Google’s AI-native IDE (distinct from Gemini CLI). Reads root AGENTS.md natively and .agent/skills/ (singular agent — not to be confused with Amp’s plural .agents/skills/).

Terminal window
npx agent-kit sync --target antigravity
{
  "scripts": { "postinstall": "agent-kit sync" },
  "agentKit": { "targets": ["antigravity"] }
}

agent-kit creates: .agent/skills/<pkg-slug>-<skill-name>/SKILL.md. No derived file.

Reload quirk: Reload the Antigravity workspace after the first sync.

Using more than one agent at once

You can pass multiple targets to a single sync invocation, and the agentKit.targets array in package.json accepts a list too. This is the right answer when a project has contributors using different agents — every supported agent picks up the same skills with one install.

Terminal window
npx agent-kit sync --target claude,cursor,codex
{
  "scripts": { "postinstall": "agent-kit sync" },
  "agentKit": { "targets": ["claude", "cursor", "codex", "copilot"] }
}

Every listed target gets its own .<tool>/skills/ directory. Skill content is copied (not symlinked) — disk overhead is small, but you can mix and match without worrying about cross-tool state. The derive step still emits CLAUDE.md, .gemini/GEMINI.md, .github/copilot-instructions.md, and CONVENTIONS.md on every sync regardless of which skill targets you pick.

Where to go next

  • Overview — what agent-kit is and how it fits together
  • CLI usage — every flag, every command, exact invocations
  • Recipes — monorepo wiring, watch mode, CI guardrails, pick/omit filtering
  • Authoring skills — for package authors who want to ship reusable skills inside their own npm package