@agentkit-js/claude-agent-sdk
Drop agentkit-js sandbox kernels into the Anthropic Claude Agent SDK
as native tools. Edge-safe code execution, one CapabilityManifest
shared with your MCP servers, no external sandbox provider.

Why this exists
Claude Agent SDK (formerly
the Anthropic agent runtime) is Anthropic's first-party way to build agents
that target Claude Code, the desktop, or any hosted Claude application. It
ships with a clean tool registration model — but no sandbox. If your model
generates code, you have to either trust it (don't), spin up a container
(latency, ops cost), or write a node:vm shim (bans you from edge
runtimes).
This package gives you a ClaudeAgentTool backed by an agentkit kernel
(QuickJSKernel, PyodideKernel, WasmtimeKernel, or RemoteSandboxKernel).
Pick a tier; pick a manifest; the tool drops straight into the agent SDK's
tools array.
Install
npm install @anthropic-ai/sdk @agentkit-js/claude-agent-sdk @agentkit-js/kernel-quickjs \
quickjs-emscripten @jitl/quickjs-wasmfile-release-sync zod
One-shot snippet evaluation
import { sandboxedJsClaudeTool } from "@agentkit-js/claude-agent-sdk";
import { QuickJSKernel } from "@agentkit-js/kernel-quickjs";
const tool = sandboxedJsClaudeTool({
kernel: new QuickJSKernel(),
capabilities: { allowedHosts: ["api.example.com"], cpuMs: 5000 },
});
Code-mode: collapse N tools
import { codeModeClaudeTool } from "@agentkit-js/claude-agent-sdk";
import { ToolRegistry } from "@agentkit-js/core";
const tools = new ToolRegistry();
const tool = codeModeClaudeTool({
kernel: new QuickJSKernel(),
tools,
capabilities: { allowedHosts: ["api.github.com"] },
});
Capability manifest
Same CapabilityManifest as every other agentkit kernel — see
docs/guides/code-mode.md.
See also