Most AI kits ship opaque SDKs. You call an API, get a response, and hope the abstraction holds up when you need to change prompts, tools, or model settings. AgentCN takes a different approach: install agents like components, with real TypeScript source you can edit and own.
Why I built it
When shipping AI agents in production, I kept hitting the same wall. Frameworks were great for demos, but customization meant fighting the SDK or forking internals. I wanted something closer to how we use UI libraries: run a command, get source in your repo, change what you need.
AgentCN is an open-source kit with a CLI, hosted registry, and agents built on the Vercel AI SDK.
Quick start
Install the web agent into any project:
npx agentcn@latest add web-agentList what’s available:
npx agentcn@latest listPreview files before installing:
npx agentcn@latest add web-agent --dry-runAfter install, agent source lands in your repo (typically under ai/agents/). Prompts, tools, and runtime are plain TypeScript you can change directly.
How it works
- Registry: Agent manifests live at agentcn.dev/r. The CLI reads them and pulls the right files into your project.
- CLI install:
npx agentcn add <agent>copies source, not a black-box dependency. - Editable source: You own the code. Swap models, tweak prompts, add tools without waiting on a library update.
- Docs + demos: Setup guides and previews live at agentcn.dev/docs.
Available agents
Run npx agentcn@latest list for the full set. Current agents include:
- Web Agent — browser automation, web search, deep research, and Q&A
- Extraction Agent — structured data extraction from documents
- Ecommerce Agent — product and catalog workflows
Each agent has its own docs page with wiring examples and required API keys.
Example: calling an agent
After installing the web agent, import from the generated path in your app:
import { webAgent } from '@/ai/agents/web';
const result = await webAgent.generate({
prompt: 'What are the latest updates on the Vercel AI SDK?',
});The exact import path depends on your project setup. Full guides are in the Web Agent docs.
Environment variables
Most agents need an Anthropic API key:
ANTHROPIC_API_KEY=your_key_hereThe web agent also needs search and browser keys:
EXA_API_KEY=your_exa_key
ANCHOR_API_KEY=your_anchor_keyWhat’s next
AgentCN is early but growing. The registry, CLI, and docs site are all open source. If you’re building with AI agents and want source you can actually change, give it a try.