This skill installs a wrapper, a streaming route and tests that run Claude agents on the Anthropic SDK’s tool runner, inside the app’s own process on any Node host. The agent has exactly the tools the wrapper gives it, with no shell and no web access built in.
An agent is a directory with an AGENTS.md as its whole system prompt, plus optional knowledge, local skills and an MCP config. It also shows how to build tools that act for the signed-in caller without ever taking an identity as input.
What You Get
- A wrapper, an MCP helper, a route that lists and streams agents, a unit test and a live test
- A script that creates an agent directory, and one that lists each MCP server’s tools next to the granted ones
- Knowledge tools confined to the agent’s knowledge folder, and a load_skill tool for agent-local skills
- A pattern for caller tools where identity is closed over, never passed as input
- Two finished example agents, copied only on request
Common Use Cases
- Adding a support agent that reads a subscriber’s invoices through app-owned tools
- Embedding a knowledge agent that answers only from local markdown files
- Giving an agent a short list of MCP tools such as web research
- Deploying an agent on a serverless host whose time limit covers the slowest run
What Makes It Unique
It is the default of two embedded-agent skills and runs with no subprocess, so a serverless function can host it. Tools are the whole boundary: nothing is built in, file paths are resolved through symlinks, exceptions reach the model as a fixed sentence, and one-way actions go through a button the app shows, never through model text.
Under The Hood
The skill is mostly instructions, with two Node scripts and TypeScript templates. The agent runs npm installs, copies template files into the app and runs new-agent.mjs, which writes a new agent directory. The discover script opens each MCP server from the agent’s config, which can reach the network with the agent’s secrets, and only reads its tool list. At run time the wrapper calls the Anthropic API with the app’s key, runs the tools the model asks for, and starts stdio MCP servers as child processes. The live tests call the real model.