Build with AI
Point your AI coding agent at Kunkun — install the skill, feed it the docs, and let it pick the right extension type and scaffold it.
Most extensions are now built with an AI agent (Claude Code, Cursor, Copilot, …). These docs are written to be read by machines as well as people — so you can hand the whole thing to your agent and let it do the work. This page shows how.
The fastest path
Install the Kunkun skill
A skill is a bundle of step-by-step instructions, references, and examples your agent loads on demand.
npx skills install kunkunsh/kunkunThis installs two skills:
kunkun-plugin-development— designing, building, and reviewing extensions: manifests, command modes, services, permissions, host APIs, resource bindings.kunkun-extension-testing— testing extensions and services without launching the desktop app.
Reload your agent
Some editors/agents only pick up newly-installed skills after a reload or restart. Verify the skill is available before you start building (ask your agent to list its skills). If it's loaded, tell your agent to follow it.
…or just paste a prompt
No skill support? Give any agent this one line — it fetches a short decision-tree guide that walks it through the whole build:
Build a Kunkun extension by following the guide at
https://docs.kunkun.sh/instruction.mdThat instruction.md file asks you what you want to build, picks the right command mode, and reads the specific docs it needs — nothing is hard-coded, so it never goes stale.
Feed the docs to your agent
The whole site is available in machine-readable form. Point your agent at these instead of pasting screenshots or guessing:
llms.txt
An index of every page with its description — a table of contents for the model.
llms-full.txt
Every page concatenated into one file. Great for smaller models with big context windows.
Any page as Markdown
Append a page path to /llms.mdx/docs/ to fetch its raw Markdown — e.g. /llms.mdx/docs/building/record-storage.
Ask AI
Every page has a "Copy Markdown" / "Open in…" button, and the floating Ask AI assistant searches the whole docs for you.
For framework docs (React, Svelte, Vue, Vite) rather than Kunkun docs, point your agent at Context7 or the library's own site — your training data may be out of date.
Which extension should the AI build?
You don't need to know upfront — the agent will ask. Under the hood it's a small decision tree:
- No UI? → a headless command (on-demand,
interval/cron, or event-triggered). - Simple UI? (list, form, detail) → a worker-view — the best default.
- Complex/custom UI? (charts, canvas, a whole SPA, a database visualizer) → a custom-view in the framework of your choice.
- Then: permissions, optional storage, backends, and services — each pulled in only if needed.
See Extension Types for the full comparison, or let instruction.md drive the conversation.
Keep humans in the loop
The agent should still show you its plan and the manifest before writing lots of code. Extensions declare real permissions and install as built artifacts — review what it asks for.