English | 中文

Harness Research MCP

Give any AI Agent professional-grade deep research capabilities.

An MCP (Model Context Protocol) server that performs real-time deep research across 5 data sources, evaluates source credibility with the CRAAP framework, cross-verifies findings, and generates professional reports — all driven by affordable models like Kimi K2.5.

A full deep research report takes ~10 minutes and can be driven by Kimi K2.5 at ~$0.01/run.

For AI Agents: How to install this tool

This is a Node.js MCP server, NOT a Python package. Do NOT git clone or pip install.

Step 1 — One-time setup (run in user's terminal):

npx harness-research-mcp setup

This launches an interactive wizard to configure API keys (stored in ~/.harness-research/.env).

Step 2 — Register as MCP server in your agent's config file:

Claude Desktop / Cursor / Windsurf — edit MCP config JSON:

{
  "mcpServers": {
    "harness-research": {
      "command": "npx",
      "args": ["-y", "harness-research-mcp"]
    }
  }
}

OpenCode — add to ~/.config/opencode/opencode.json:

{
  "mcp": {
    "harness-research": {
      "command": "npx",
      "args": ["-y", "harness-research-mcp"]
    }
  }
}

That's it. After registration, the agent gets 3 tools: harness_research (full report), harness_search (quick search), harness_status (progress check).

Do NOT clone this repo, build from source, or run node/python directly. npx handles everything.

Async workflow — no timeout issues: harness_research returns a task_id immediately (< 1 second). The research runs in the background (~10 minutes). You must poll harness_status(task_id) every 30–60 seconds until status is "completed" or "failed". This design means the default 60-second MCP timeout will NOT be a problem.

If your MCP client has a configurable timeout, setting it to 1200 seconds (20 minutes) is still recommended as a safety margin:

{
  "mcpServers": {
    "harness-research": {
      "command": "npx",
      "args": ["-y", "harness-research-mcp"],
      "timeout": 1200
    }
  }
}

Why This Tool Exists

The Problem with Current "Deep Research" Tools

Existing deep research tools (Perplexity Deep Research, ChatGPT Research, Gemini Deep Research, etc.) share fundamental flaws:

How Harness Research Is Different

Core principle: The LLM only "thinks" — it never "knows." All factual data comes from real-time search.

Related MCP server: GPT Researcher MCP Server

The 6-Step Research Pipeline

User: "Research the global AI chip market landscape in 2025"
         │
         ▼
Step 1 ── Research Plan (LLM)
         │  Generate chapter structure + search keywords
         ▼
Step 2 ── 5-Source Parallel Search (Code)
         │  Tavily + Brave + arXiv + PubMed + Tushare
         │  Dedup → cap at 50 results
         ▼
Step 3 ── CRAAP Source Evaluation (Code + LLM)
         │  Code pre-filter: T5 eliminated, >3yr eliminated
         │  LLM batch scoring: Relevance + Accuracy + Purpose
         │  Weighted average → filter low-scoring sources
         ▼
Step 4 ── Cross-Verification (LLM)
         │  Data triangulation + conflict detection + counterintuitive findings
         ▼
Step 5 ── Parallel Writing (LLM)
         │  All chapters in parallel + executive summary
         ▼
Step 6 ── Render Output (Code)
         │  HTML + DOCX + PDF (macOS) + Markdown
         ▼
    Professional research report (~10 minutes)

Quick Start

1. Setup (one-time)

npx harness-research-mcp setup

The interactive wizard will guide you through:

• Configuring search API keys (Tavily or Brave, at least one)

• Configuring an LLM API key (Kimi K2.5 recommended — cheapest option)

• Optional: Tushare (Chinese financial data), NCBI (PubMed academic search)

• Automatic API connectivity test

2. Register with Your AI Agent

Copy the appropriate config for your Agent framework:

Claude Desktop / Cursor / Windsurf:

{
  "mcpServers": {
    "harness-research": {
      "command": "npx",
      "args": ["-y", "harness-research-mcp"]
    }
  }
}

OpenClaw:

openclaw mcp set harness-research '{"command":"npx","args":["-y","harness-research-mcp"]}'

OpenCode:

// ~/.config/opencode/opencode.json
{
  "mcp": {
    "harness-research": {
      "command": "npx",
      "args": ["-y", "harness-research-mcp"]
    }
  }
}

3. Use It

Just tell your Agent:

"Do a deep research on the global AI chip market landscape in 2025"

The Agent will automatically call harness_research and return the full report in ~10 minutes.

Three MCP Tools

API Keys Explained

Why Do You Need These Keys?

Harness Research does not rely on any LLM's historical knowledge. All information is fetched in real-time from the internet. This requires calling various search and data APIs.

Minimum: 1 search key + 1 LLM key = 2 keys to get started.

Why Kimi K2.5?

• Cost: ~$0.01 per full research session (vs. GPT-4 at $1-5)

• Chinese support: Native Chinese language, no translation layer needed

• Context: 128K token window — handles large volumes of search results

• Reliability: 99.9%+ API availability

Output Formats

CRAAP Evaluation Framework

Every source is scored across 5 dimensions:

6-Tier Source Classification

Built-in 530+ domain credibility database covering major governments, academia, media, and professional institutions worldwide.

Diagnostics

npx harness-research-mcp doctor

Architecture

┌──────────────────────────────────────────┐
│  Claude / Cursor / OpenClaw / OpenCode   │
│            (MCP Client)                  │
└────────────────┬─────────────────────────┘
                 │ stdio (MCP Protocol)
                 ▼
┌──────────────────────────────────────────┐
│      harness-research-mcp (Node.js)      │
│                                          │
│  Tools:                                  │
│    harness_research — full deep research  │
│    harness_search   — quick multi-search  │
│    harness_status   — progress query      │
│                                          │
│  6-Step Pipeline:                         │
│    Plan → Search → CRAAP → Verify →       │
│    Write → Render                         │
│                                          │
│  Pure Node.js. Zero Python dependency.    │
└──────────────────────────────────────────┘

Development

git clone https://github.com/Nimo1987/harness-research.git
cd harness-research
npm install
npm run build

License

Apache 2.0