MCP Server

What telleroutlook/agentkit-js/mcp-server does

telleroutlook/agentkit-js/mcp-server exposes an agent or tool registry through Model Context Protocol while reducing the host-visible interface to two tools: docs_search and execute_code. Instead of publishing every downstream tool directly, the model submits a script through execute_code; that script can call the registered tools it needs with callTool(...).

This design is intended for agents with many tools. The repository reports that, with 30 downstream tools, the bootstrap-token cost is 13.6% of the equivalent direct-tool arrangement. The server is beta software, so its API may change between minor releases.

How it works

A ToolRegistry holds the downstream tools. createCodeModeServer() combines that registry with a selected execution kernel and a CapabilityManifest, producing an MCP server with the two-tool surface. The manifest is applied at the kernel boundary to code executed through execute_code, including code that invokes downstream tools.

The manifest can define permitted network hosts, filesystem paths, CPU time, and memory. An empty allowedHosts list disables outbound network access, while an empty allowedPaths list disables filesystem access. CPU and memory limits can be set with cpuMs and memoryLimitBytes.

Execution can use an in-process node:vm kernel, WASM-based QuickJS, Pyodide, or Wasmtime kernels, or a remote microVM backed by E2B or Cloudflare Sandbox. The same capability model is used across these kernel choices.

Setup and configuration

Install the package with npm and create a small Node.js entry point. Register the tools in a ToolRegistry, choose a kernel such as QuickJSKernel, and pass both to createCodeModeServer() along with the capability settings and server metadata.

For local MCP clients, call runStdio() with the created server. The package also provides createFetchHandler() for HTTP or streamable deployments such as Cloudflare Workers and Vercel. Applications with their own transport can call the server’s asynchronous handle(req) method directly.

Running the package with npx starts a zero-configuration stdio server using a VM kernel and no downstream tools. That mode is useful for basic sanity checks or introspection rather than a configured agent deployment.

Tools and capabilities

• docs_search provides the documentation-search side of the code-mode interface.
• execute_code runs model-supplied code and lets it dispatch to registered tools.
• ToolRegistry supplies the downstream tools available to executed code.
• CapabilityManifest applies host, path, CPU, and memory policy.
• runStdio(), createFetchHandler(), and handle() support different transport arrangements.

Limitations and notes

The project is marked beta, and its API shape can change in minor releases. A code-mode server with no registered downstream tools does not provide useful agent actions. Resource and access policy must be configured deliberately; for example, an empty host list prevents outbound network requests. The server’s token-saving behavior is most relevant when the registry contains many downstream tools.