xgen-harness-engine-node

TypeScript engine for xgen-harness — reads spec.json (compiled from a Python harness workflow) and runs the same 13-stage pipeline. Fully equivalent — every stage setting from the original workflow is preserved.

Install

npm install xgen-harness-engine-node

Use as MCP server

Wrapper packages (e.g. xgen-harness-my_agent) bundle a spec.json plus a 1-line bin/cli.js that calls this engine:

const { serveMcp } = require("xgen-harness-engine-node");
const spec = require("./spec.json");
serveMcp(spec);

The wrapper is the unit Claude Desktop / Cursor / mcp-station picks up:

{
  "mcpServers": {
    "my-agent": {
      "command": "npx",
      "args": ["-y", "xgen-harness-my_agent"]
    }
  }
}

Programmatic use

import { runOnce, loadSpec } from "xgen-harness-engine-node";

const spec = loadSpec("./spec.json");
const result = await runOnce(spec, "What is 2+2?");
console.log(result.output);

What's inside

• 13 pipeline stages (s00_harness ... s11_finalize) ported 1:1 from the Python xgen-harness engine
• 4 LLM providers — Anthropic, OpenAI, vLLM (with Qwen native <tool_call> XML parser), Bedrock-ready
• 4 builtin Policy Guards — cost_cap / max_loop / pii_block / domain_allow
• Tool dispatch — http / mcp_session / rag (frozen at publish time, no Python NodeClass dependency)

Spec source

A spec.json is produced by Python:

from xgen_harness.compile import compile_workflow_to_npm
result = compile_workflow_to_npm(
    harness_config=config,
    gallery_name="my_agent",
    gallery_version="0.1.0",
    out_dir="./dist",
)

This produces a complete npm tarball (xgen-harness-my_agent-0.1.0.tgz) with the wrapper + spec + cli.

License

UNLICENSED (private use). Contact the maintainer for distribution rights.