Skip to main content
Glama
research-mcp
by KB01111

research-mcp

One MCP server + OpenAI plugin that searches 9 sources, scrapes full content, reranks, and synthesizes a cited answer with a free LLM.

Sources (9)

Source

Default?

Auth

Notes

Firecrawl (your fork)

✅

FIRECRAWL_API_URL + optional FIRECRAWL_API_KEY

Calls POST /v1/search

GitHub Trending

✅

none

HTML scrape, no key

arXiv

✅

none

Official arxiv lib

Tavily

✅

TAVILY_API_KEY

Primary fallback scraper, 1k/mo free

OSS Insight

✅

none

Trending repos, free API

dev.to

✅

none

Public API

libhunt

❌ opt-in

none

HTML scrape

X.com (Twitter)

❌ opt-in

TWITTERAPI_IO_KEY

No free public X API

Google SERP

❌ opt-in

SERPER_API_KEY or SERPAPI_API_KEY

No free public SERP API

Related MCP server: openresearch-mcp

Free LLM & reranker

  • Summarizer: Groq (llama-3.3-70b-versatile, free) → Gemini 2.0 Flash → OpenRouter free → extractive fallback

  • Reranker: Cohere rerank-english-v3.0 (1k/mo free) → Jina v2 → local term-overlap

  • Scraper: Firecrawl → Tavily extract → Jina Reader (r.jina.ai, no key)

Install

cd research-mcp
pip install -e .
cp .env.example .env
# fill in whichever keys you have; the rest are optional

Run the MCP (stdio, for Claude Desktop / goose)

python -m mcp_server

Add to your Claude Desktop config:

{
  "mcpServers": {
    "research": {
      "command": "python",
      "args": ["-m", "mcp_server"],
      "cwd": "C:\\Users\\kevin\\research-mcp"
    }
  }
}

Run the OpenAI plugin shim (HTTPS)

python -m openai_plugin.app
# then serve over HTTPS via Caddy/Nginx and point ChatGPT at
# https://your.host/.well-known/openai-plugin.json

Tools exposed

Tool

Purpose

research(query, sources?, max_results?, style?)

End-to-end: search → scrape → rerank → summarize

search_all(query, sources?, max_results?)

Search only, no LLM

scrape_url(url, method?)

Single URL scrape

rerank_docs(query, docs, model?)

Rerank a list of strings

summarize_docs(docs, query, style?, model?)

Final synthesis

Smoke test (no keys required)

python tests/smoke.py

Runs arxiv → local rerank → extractive summary. Proves the pipeline is wired correctly without spending any free-tier quota.

Docker

docker compose --profile mcp up --build       # MCP over stdio
docker compose --profile plugin up --build    # OpenAI plugin on :8000

Layout

mcp_server/
  server.py                # fastmcp entry
  config.py                # env-driven settings
  adapters/                # one file per source, normalized SearchResult
  tools/
    search.py              # fan-out + dedupe
    scrape.py              # Firecrawl / Tavily / Jina
    rerank.py              # Cohere / Jina / local
    summarize.py           # Groq / Gemini / OpenRouter
openai_plugin/
  app.py                   # FastAPI shim re-exporting the same tools
  manifest.json
tests/smoke.py
Dockerfile, docker-compose.yml

Available Tools

5 tools
rerank_docsA

Rerank a list of document strings by relevance to the query.

model: "cohere" | "jina" | "local" Returns: [{index, relevance_score}, ...] sorted desc.

Inputs: docs (required), query (required), model. Returns: result

researchA

End-to-end: search all sources -> scrape -> rerank -> summarize.

Inputs: query (required), style, sources, max_results. Returns: result

scrape_urlC

Scrape a single URL and return markdown content.

method: "firecrawl" | "tavily" | "jina"

Inputs: url (required), method. Returns: result

search_allC

Fan out the query to every source and return deduped SearchResult dicts.

Inputs: query (required), sources, max_results. Returns: result

summarize_docsB

Synthesize a final answer from a list of document strings.

style: "detailed" | "bullets" | "tldr" model: "groq" | "gemini" | "openrouter"

Inputs: docs (required), query (required), model, style. Returns: result

Tool Schema Changelog

Recent tool additions, removals, and schema changes observed during successful MCP inspections.

Loading tool schema changelog…

    TDQS

    A3.6/5.0

    Scored across 5 tools

    Disambiguation5/5

    Each tool targets a distinct stage in the research pipeline (search, scrape, rerank, summarize), and `research` serves as an orchestrator that combines these steps. No two tools have overlapping core functions, so an agent can easily select the right tool for a specific need.

    Naming Consistency4/5

    Most tools follow a consistent verb_noun pattern (scrape_url, rerank_docs, summarize_docs), but `research` and `search_all` deviate slightly. This is a minor inconsistency that does not hinder readability.

    Tool Count5/5

    With 5 tools, the server is well-scoped for a research workflow. Each tool covers a necessary step in the pipeline, and the count is neither too sparse nor bloated.

    Completeness5/5

    The tool set provides end-to-end coverage from searching multiple sources to summarizing final answers. The individual building blocks (search_all, scrape_url, rerank_docs, summarize_docs) and the composite `research` tool leave no obvious gaps for typical research tasks.

    Related MCP Connectors

    Related MCP Servers

    • A
      license
      Not graded
      quality
      D
      maintenance
      MCP server for internet search via direct Google and DuckDuckGo HTML scraping with AI-powered result normalization and optional summarization, requiring no API keys for search.
      MIT
    • A
      license
      Not graded
      quality
      C
      maintenance
      Zero-auth multi-source research MCP server that enables web search, reading URLs, PDFs, GitHub repos, and querying Hacker News, Stack Overflow, Semantic Scholar, and YouTube transcripts without API keys.
      58 PyPI
      12
      Apache 2.0