mcp-research-powerpack

MCP server that gives your AI assistant research tools. Google search, Reddit deep-dives, web scraping with LLM extraction, and multi-model deep research — all as MCP tools that chain into each other.

npx mcp-research-powerpack

five tools, zero config to start. each API key you add unlocks more capabilities.

tools

tools are designed to chain: web_search suggests calling scrape_links, which suggests search_reddit, which suggests get_reddit_post, which suggests deep_research for synthesis.

install

Claude Desktop / Claude Code

add to your MCP config:

{
  "mcpServers": {
    "research-powerpack": {
      "command": "npx",
      "args": ["mcp-research-powerpack"],
      "env": {
        "SERPER_API_KEY": "...",
        "OPENROUTER_API_KEY": "..."
      }
    }
  }
}

from source

git clone https://github.com/yigitkonur/mcp-research-powerpack.git
cd mcp-research-powerpack
pnpm install && pnpm build
pnpm start

HTTP mode

MCP_TRANSPORT=http MCP_PORT=3000 npx mcp-research-powerpack

exposes /mcp (POST/GET/DELETE with session headers) and /health.

API keys

each key unlocks a capability. missing keys silently disable their tools — the server never crashes.

configuration

optional tuning via environment variables:

how it works

search ranking

results from multiple queries are deduplicated by normalized URL and scored using CTR-weighted position values (position 1 = 100.0, position 10 = 12.56). URLs appearing across multiple queries get a consensus marker. threshold tries >= 3, falls back to >= 2, then >= 1.

Reddit comment budget

global budget of 1,000 comments, max 200 per post. after the first pass, surplus from posts with fewer comments is redistributed to truncated posts in a second fetch pass.

scraping pipeline

three-mode fallback per URL: basic → JS rendering → JS + US geo-targeting. results go through HTML-to-markdown conversion (turndown), then optional LLM extraction with a 100k char input cap and 8,000 token output per URL.

deep research

32,000 token budget divided across questions (1 question = 32k, 10 questions = 3.2k each). Gemini models get google_search tool access. Grok/Perplexity get search_parameters with citations. primary model fails → automatic fallback.

file attachments

deep_research can read local files and include them as context. files over 600 lines are smart-truncated (first 500 + last 100 lines). line numbers preserved.

concurrency

all clients use manual retry with exponential backoff and jitter. the OpenAI SDK's built-in retry is disabled (maxRetries: 0).

project structure

src/
  index.ts                — entry point, STDIO + HTTP transport, signal handling
  worker.ts               — Cloudflare Workers entry (Durable Objects)
  config/
    index.ts              — env parsing (lazy Proxy objects), capability detection
    loader.ts             — YAML → Zod → JSON Schema pipeline, cached
    yaml/tools.yaml       — single source of truth for all tool definitions
  schemas/
    deep-research.ts      — Zod validation for research questions + file attachments
    scrape-links.ts       — Zod validation for URLs, timeout, LLM options
    web-search.ts         — Zod validation for keyword arrays
  tools/
    registry.ts           — tool lookup → capability check → validate → execute
    search.ts             — web_search handler
    reddit.ts             — search_reddit + get_reddit_post handlers
    scrape.ts             — scrape_links handler
    research.ts           — deep_research handler
  clients/
    search.ts             — Serper API client
    reddit.ts             — Reddit OAuth + comment fetching
    scraper.ts            — scrape.do client with fallback modes
    research.ts           — OpenRouter client with model-specific handling
  services/
    llm-processor.ts      — shared LLM extraction (singleton OpenAI client)
    markdown-cleaner.ts   — HTML → markdown via turndown
    file-attachment.ts    — local file reading with line ranges
  utils/
    concurrency.ts        — bounded parallel execution (pMap, pMapSettled)
    url-aggregator.ts     — CTR-weighted scoring and consensus detection
    errors.ts             — error classification, fetchWithTimeout
    logger.ts             — MCP logging protocol
    response.ts           — standardized output formatting

deploy

Cloudflare Workers

npx wrangler deploy

uses Durable Objects with SQLite storage. YAML-based tool definitions are replaced with inline definitions in the worker entry since there's no filesystem.

license

MIT