Using an agent

Add

npx agentskit add <agent>

Copies the agent's source into ./agents/<agent>/ (override with --out). You own the code.

npx agentskit add legal-contract-reviewer
npm install @agentskit/core @agentskit/runtime @agentskit/adapters

import { openai } from '@agentskit/adapters'
import { createLegalContractReviewerAgent } from './agents/legal-contract-reviewer/agent'

const agent = createLegalContractReviewerAgent({
  adapter: openai({ apiKey: process.env.OPENAI_API_KEY!, model: 'gpt-4o' }),
})
const { content } = await agent.run('Review this NDA: …')

Swap the adapter for any provider (anthropic, gemini, ollama, …) — no lock-in.

Compare before installing

The agent catalog can narrow the registry to independently reviewed or directly runnable agents. Results can be ordered by recommendation, independent review score, or name.

Each category also has a stable, indexable URL, such as Coding agents or Legal agents. Category pages list every matching agent and link to its canonical detail page.

Select two or three agents and choose Compare agents to inspect review evidence, packages, required environment variables, capabilities, and install commands side by side. Filter and comparison state stays in the URL, so the resulting view can be shared with a teammate.

Catalog URLs accept q, category, reviewed=1, runnable=1, sort, page, and compare query parameters. Comparison pages are intentionally excluded from search indexing because each URL represents a temporary decision workspace rather than canonical agent content.

Run it in one command

npx agentskit add <agent> --run "<task>" --provider ollama

Copies and runs the agent (its system prompt, as data — never executing the copied code). Use --provider/--model/--api-key, or a local model with --provider ollama (no key).

Keep it up to date

npx agentskit diff <agent>     # show how your copy differs from the registry
npx agentskit update <agent>   # apply the registry version

Compose, ground, secure

Every agent's factory accepts optional config — all overridable:

createXAgent({
  adapter,
  tools,       // tools / integrations / MCP (toolsFromMcpClient)
  memory,      // conversation context
  retriever,   // RAG grounding
  delegates,   // sub-agents (orchestration)
  onConfirm,   // per-tool permission gate (HITL / RBAC)
  observers,   // tracing / audit
})

Agents also expose .asHandle() for multi-agent topologies (supervisor, swarm, …).